CN111368408A - Seq2 seq-based data center energy efficiency optimization continuous decision method - Google Patents

Seq2 seq-based data center energy efficiency optimization continuous decision method Download PDF

Info

Publication number
CN111368408A
CN111368408A CN202010122279.6A CN202010122279A CN111368408A CN 111368408 A CN111368408 A CN 111368408A CN 202010122279 A CN202010122279 A CN 202010122279A CN 111368408 A CN111368408 A CN 111368408A
Authority
CN
China
Prior art keywords
model
encoder
sequence
heating
decoder
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010122279.6A
Other languages
Chinese (zh)
Other versions
CN111368408B (en
Inventor
张发恩
马凡贺
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ainnovation Nanjing Technology Co ltd
Original Assignee
Ainnovation Nanjing Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Ainnovation Nanjing Technology Co ltd filed Critical Ainnovation Nanjing Technology Co ltd
Priority to CN202310735629.XA priority Critical patent/CN116611347A/en
Priority to CN202010122279.6A priority patent/CN111368408B/en
Publication of CN111368408A publication Critical patent/CN111368408A/en
Application granted granted Critical
Publication of CN111368408B publication Critical patent/CN111368408B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2113/00—Details relating to the application field
    • G06F2113/08—Fluids
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/06—Power analysis or power optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/08—Thermal analysis or thermal optimisation
    • 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)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Geometry (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Air Conditioning Control Device (AREA)

Abstract

The invention discloses a method for continuously deciding energy efficiency optimization of a data center based on seq2seq, which comprises the following steps: step S1, establishing a time sequence model of the water circulation of the heating and ventilation system based on the seq2seq structure of the recurrent neural network; step S2, evaluating the performance of the time sequence model; and step S3, carrying out constrained continuous operation strategy optimization on the whole situation of the heating and ventilation system by utilizing the time sequence model with the performance meeting the expectation. Aiming at the characteristics of the heating and ventilation system, the autocorrelation time sequence characteristics of the heating and ventilation system are taken as observable hidden variables and added into an LSTM-Cell model training network, and a time sequence model formed by training can carry out constrained continuous operation strategy optimization on the whole heating and ventilation system, so that the technical problems that the future energy consumption condition of the heating and ventilation system cannot be predicted by various energy efficiency optimization methods at present and the optimal decision cannot be made on the whole heating and ventilation system in a continuous time step are solved, and the energy saving and emission reduction effects of a data center are greatly improved.

Description

Seq2 seq-based data center energy efficiency optimization continuous decision method
Technical Field
The invention relates to the field of data mining and machine learning, in particular to a seq2 seq-based continuous decision method for optimizing energy efficiency of a data center.
Background
In recent years, with the development of technologies such as cloud service, big data and AI calculation, enterprises and governments have built a large number of data centers, but the energy consumption of the data centers is generally high, and the average PUE value (index for evaluating the energy efficiency of the data centers) is 2.2-3.0. According to statistics of relevant departments, the electricity consumption of a data center in China accounts for about 3% of the total electricity consumption of the whole society, and the proportion is increasing year by year.
At present, there are many researches on energy conservation and emission reduction of a data center, and existing or newly developed energy consumption simulation software is usually adopted to simulate the energy consumption condition of the data center so as to assist the design decision and energy efficiency optimization of the data center. However, most of the researches are biased to research the energy saving potential of the data center in the design stage, and the energy efficiency optimization problem after the data center is put into use is not considered. For example, 2016 google engineers provide a deep learning model prediction control method, and the method helps the heating and ventilation engineers to better control and optimize the heating and ventilation system by finding the relationship between the control point of the heating and ventilation system and the PUE in the data center. However, the existing energy efficiency optimization methods have the following two defects:
1. the time sequence relation of water circulation of the heating and ventilation system is not considered, the current energy consumption situation of the heating and ventilation system can be predicted only according to the current operating environment of the heating and ventilation system, and the future energy consumption situation of the heating and ventilation system cannot be predicted and optimized in a decision mode;
2. the local decision optimization can be performed on the heating and ventilation system only in a single time step, the optimal decision on the whole heating and ventilation system cannot be performed in continuous time steps, and the energy-saving and emission-reducing effects on the data center are limited.
Disclosure of Invention
The invention aims to provide a method for continuously deciding the energy efficiency optimization of a data center of seq2seq, so as to solve the technical problem.
In order to achieve the purpose, the invention adopts the following technical scheme:
the method for continuously deciding the energy efficiency optimization of the data center based on the seq2seq comprises the following steps:
step S1, establishing a time sequence model of the water circulation of the heating and ventilation system based on the seq2seq structure of the recurrent neural network;
step S2, evaluating the performance of the time sequence model;
and step S3, carrying out constrained continuous operation strategy optimization on the whole situation of the heating and ventilation system by utilizing the time sequence model with the performance meeting the expectation.
As a preferable embodiment of the present invention, the constructing the time series model in step S1 specifically includes the following steps:
step S11, building a multilayer Encoder model by using an OH-LSTM Cell neural network structure;
step S12, setting the input sequence of the Encoder model as (X + OH, M) and the output sequence of the Encoder model as (y1+ OH, M);
x is used for representing historical time sequence information of the heating and ventilating system in a specified time period;
OH is used for representing an observable hidden variable of the heating and ventilation system in the specified time period;
x + OH is the combined characteristic of the historical time sequence information and characteristic information combination input into the Encoder model in the specified time period;
y1 is used to represent the prediction target of the Encoder model;
y1+ OH represents that the output of the Encoder model includes the predicted target y1 and the observable hidden variable OH;
m is used for representing the step number of the time step of the input sequence or the output sequence of the Encoder model;
step S13, setting a loss function of the Encoder model;
step S14, constructing the training data of the Encoder model according to the sequence structure set in the step S12, and then training to form the Encoder model by taking the training data as a model training sample;
s15, cutting an input sequence of the Encoder model into (X + OH, M), and cutting an output sequence of the Encoder model into (y1+ OH, 1);
"1" represents a single said time step that the Encoder model last outputs;
step S16, constructing a multilayer Decoder model by using the OH-LSTM Cell neural network structure;
step S17, setting the input sequence of the Decode model as (X + OH, M) and the output sequence of the Decode model as (y2+ OH, N);
n is used for representing the step number of the time step output by the Decoder model;
y2 is used to represent the prediction target of the Decoder model;
y2+ OH denotes that the output of the Decoder model includes the predicted target y2 and the observable hidden variable OH;
step S18, setting a loss function of the Decoder model;
step S19, constructing training data of an Encoder-Decoder model according to the sequence structure set in the step S12 and the step S17;
step S20, transferring the model parameters of the Encoder model to the Decode model, and training to form the Encoder-Decode model as the timing sequence model by taking the training data constructed in the step S19 as a training sample;
step S21, cutting the input sequence of the Encoder-Decoder model into (X + OH, M + N), cutting the output sequence of the Encoder-Decoder model into (y2, N),
m + N is used to represent that the input sequence of the Encoder-Decoder model contains M + N time steps.
As a preferable aspect of the present invention, the feature information includes at least the observable hidden variable.
As a preferable embodiment of the present invention, in the step S13, a loss function of the Encoder model is set to perform a weighted average operation on an output sequence loss of the Encoder model.
As a preferable embodiment of the present invention, in the step S18, a loss function of the Decoder model is set to perform an average operation on an output sequence loss of the Decoder model.
As a preferable aspect of the present invention, in step S2, the method for evaluating the performance of the time series model includes a continuous sensitivity curve analysis and evaluation method, and the continuous sensitivity curve analysis and evaluation method specifically includes the following steps:
step L1, randomly extracting a sample list of specified continuous step sizes as the input of the Encoder model;
step L2, selecting n sensitivity parameters to be analyzed and evaluated;
step L3, constructing a parameter integer list of the sensitivity parameters;
a step L4 of cartesian product combining the parameter integer list with the input sequence of the Encoder model to which the t-th time step within the consecutive step sizes specified by the step L1 is input, to construct n sets of input sequences of the Decoder model;
l5, carrying out Cartesian product combination on the input sequence of the Encoder model and the input sequences of n groups of the Decoder model, and constructing n groups of input sequences of the Encoder-Decoder model;
a step L6 of taking the input sequence of the n sets of Encoder-Decoder models constructed in the step L5 as the input of the Encoder-Decoder models and obtaining the output of the Encoder-Decoder models;
and L7, drawing a time sequence distribution curve which can represent the energy consumption situation of the heating and ventilation system in the specified time period of the continuous step length according to the output of the Encoder-Decoder model.
As a preferable aspect of the present invention, in step S2, the method for evaluating the performance of the time sequence model includes a delay sensitivity curve analysis and evaluation method, and the delay sensitivity curve analysis and evaluation method specifically includes the following steps:
step M1, randomly extracting M sections of sample list groups with appointed continuous step length as the input of the Encoder model;
step M2, selecting n sensitivity parameters to be analyzed and evaluated;
step M3, constructing a parameter integer list of the sensitivity parameters;
step M4, carrying out Cartesian product combination on the parameter integer list and M groups of input sequences of the Encoder model to construct n x M groups of input sequences of the Decoder model;
a step M5 of inputting the sensitivity parameters of each group in the n x M groups and the t-th time step in the specified continuous step into the input sequence of the Encoder model to carry out Cartesian product combination so as to construct n x M groups of input sequences of the Encoder-Decoder model;
a step M6 of taking the input sequence of the Encoder-Decoder model of the n x M groups constructed in the step M5 as the input of the Encoder-Decoder model and obtaining the output of the Encoder-Decoder model;
and step M7, drawing a time sequence distribution curve which can represent the energy consumption condition of the heating and ventilation system in different time periods according to the output of the Encoder-Decoder model.
As a preferable aspect of the present invention, the number of steps of the consecutive steps randomly extracted is 30 time steps.
In a preferred embodiment of the present invention, t is 30.
As a preferred embodiment of the present invention, in the step S3, the process of performing constrained global continuous optimization on the operation strategy of the heating and ventilation system according to the timing model includes the following steps:
step N1, extracting historical sample data of the heating and ventilation system at the current moment to construct a control parameter combination candidate set of the Decoder model in j continuous historical time steps, wherein j is a natural number more than or equal to 1;
step N2, constructing k groups of control parameter combination candidate sets;
step N3, combining k groups of control parameter combination candidate sets with environmental parameters to construct m groups of input sequences of the Decoder model;
step N4, combining the constructed input sequences of k groups of the Decoder model with the input sequences of the Encoder model which is also k groups, and constructing the input sequences of k groups of the Encoder-Decoder model;
a step N5 of inputting the input sequence of the k sets of the Encoder-Decoder models constructed in the step N4 into the Encoder-Decoder models and outputting a prediction result of the output power of the heating and ventilation system;
step N6, according to the prediction result, eliminating the control parameter combination candidate set which does not meet the constraint condition;
step N7, selecting the control parameter combination candidate set with the lowest predicted output power of the heating and ventilation system from at least one group of the control parameter combination candidate sets which are in accordance with the constraint condition as the control parameter set of the heating and ventilation system at the next time step of the current time, and issuing the control parameter set to corresponding equipment in the heating and ventilation system, wherein each equipment in the heating and ventilation system operates according to the issued control parameter;
and step N8, repeating the steps N1-N7 in the next time step of the current time so as to make continuous optimization decision on the global operation state of the heating and ventilation system.
Aiming at the characteristics of the heating and ventilation system, the autocorrelation time sequence characteristics of the heating and ventilation system are taken as observable hidden variables and added into an LSTM-Cell model training network, and a time sequence model formed by training can carry out constrained continuous operation strategy optimization on the whole heating and ventilation system, so that the technical problems that the future energy consumption condition of the heating and ventilation system cannot be predicted by various energy efficiency optimization methods at present and the optimal decision cannot be made on the whole heating and ventilation system in a continuous time step are solved, and the energy saving and emission reduction effects of a data center are greatly improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required to be used in the embodiments of the present invention will be briefly described below. It is obvious that the drawings described below are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort.
FIG. 1 is a diagram illustrating method steps of a seq2 seq-based data center energy efficiency optimization continuous decision method according to an embodiment of the present invention;
FIG. 2 is a diagram of method steps for constructing the timing model;
FIG. 3 is a diagram of the method steps for evaluating the time series model using the continuous sensitivity curve analysis evaluation method;
FIG. 4 is a diagram of the method steps for evaluating the timing model using the delay sensitivity curve analysis evaluation method;
FIG. 5 is a diagram of method steps for constrained global continuous optimization of operating strategy of the heating and ventilation system according to the timing model;
FIG. 6 is a schematic diagram of the improved OH-LSTM Cell neural network structure of the present invention;
FIG. 7 is a schematic structural diagram of the Encoder-Decoder model;
FIG. 8 is a first graphical illustration of a time series profile generated by the continuous sensitivity curve analysis and evaluation method;
FIG. 9 is a second graphical illustration of a time series profile generated by the continuous sensitivity curve analysis and evaluation method;
fig. 10 is a schematic diagram of a timing distribution curve generated by the delay sensitivity curve analysis and evaluation method.
Detailed Description
The technical scheme of the invention is further explained by the specific implementation mode in combination with the attached drawings.
Wherein the showings are for the purpose of illustration only and are shown by way of illustration only and not in actual form, and are not to be construed as limiting the present patent; to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product; it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The same or similar reference numerals in the drawings of the embodiments of the present invention correspond to the same or similar components; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "inner", "outer", etc. are used for indicating the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, it is only for convenience of description and simplification of description, but it is not indicated or implied that the referred device or element must have a specific orientation, be constructed in a specific orientation and be operated, and therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and are not to be construed as limitations of the present patent, and the specific meanings of the terms may be understood by those skilled in the art according to specific situations.
In the description of the present invention, unless otherwise explicitly specified or limited, the term "connected" or the like, if appearing to indicate a connection relationship between the components, is to be understood broadly, for example, as being fixed or detachable or integral; can be mechanically or electrically connected; they may be directly connected or indirectly connected through intervening media, or may be connected through one or more other components or may be in an interactive relationship with one another. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.
Aiming at the characteristics of the heating and ventilation system, the autocorrelation time sequence characteristics of the heating and ventilation system are taken as observable hidden variables and added into an LSTM-Cell model training network, and a time sequence model formed by training can carry out constrained continuous operation strategy optimization on the whole heating and ventilation system, so that the technical problems that the future energy consumption condition of the heating and ventilation system cannot be predicted by various energy efficiency optimization methods at present and the optimal decision cannot be made on the whole heating and ventilation system in a continuous time step are solved, and the energy saving and emission reduction effects of a data center are greatly improved.
Referring to fig. 1, the embodiment of the invention provides a method for continuously deciding energy efficiency optimization of a data center based on a seq2seq, which includes the following steps:
step S1, establishing a time sequence model of the water circulation of the heating and ventilation system based on the seq2seq structure of the recurrent neural network;
step S2, performing performance evaluation on the time sequence model;
and step S3, carrying out constrained continuous operation strategy optimization on the whole situation of the heating and ventilation system by utilizing the time sequence model with the performance meeting the expectation.
Referring to fig. 2, in step S1, the method for constructing the timing model specifically includes the following steps:
s11, building a multilayer Encoder model by using an OH-LSTM Cell neural network structure, specifically building the Encoder model by using a keras framework of the OH-LSTM Cell neural network; the OH-LSTM Cell neural network structure is a time-cycle neural network structure improved based on the existing LSTM (long-short term memory network) neural network. The improvement points of the invention are that: adding pepole of OH (observed hidden variable) into LSTM, so that the output y of the time sequence model can snoop the closed-loop information through pepole and feed back to hidden layer states (hidden states). OH is an observable hidden variable of the water circulation of the heating and ventilation system (such as the temperature and the pressure of inlet and outlet water and related variables which have indirect influence on the output power of the heating and ventilation system). The peeholie is to establish a calculation channel for an autoregressive variable, that is, variables (such as inlet and outlet water temperature, pressure and the like) related to the output power of the heating and ventilation system are predicted, and then a power value to be predicted is finally obtained according to the variable values of the autoregressive variable. OH is the autoregressive variable.
Please refer to fig. 6 for a schematic diagram of the OH-LSTM Cell neural network structure. In fig. 6, Xt represents a parameter vector of the heating and ventilation system (a vector formed by the heating and ventilation system control parameters) at the current time point;
OHt, an observable hidden variable of the heating and ventilation system at the current time point;
ht represents a hidden variable of the current time;
OHHt is the hidden layer output after OHt and Ht fusion (splicing);
c is the memory neuron of LSTM neural network.
FIG. 7 shows a schematic structural diagram of an Encoder-Decoder model according to the present invention. With continuing reference to fig. 2 and with further reference to fig. 7, after the Encoder model construction of step S11 is completed, the method proceeds to step S12,
step S12, setting the input sequence of the Encoder model as (X + OH, M) and the output sequence of the Encoder model as (y1+ OH, M);
x is used for representing historical time sequence information of the heating and ventilation system in a certain period of time;
OH is used for representing an observable hidden variable of the heating and ventilation system in the specified time period;
x + OH is the combined characteristic of the combination of the historical time sequence information and the characteristic information of the input Encoder model in the specified time period; the characteristic information at least comprises an observable hidden variable OH, and in addition, the characteristic information generally also comprises the working frequency, the external temperature and humidity and other characteristics of equipment such as a heating and ventilation system cooling tower, a cooling pump, a cold machine and the like;
y1 is used to represent the prediction target of the Encoder model;
y1+ OH represents that the output of the Encoder model contains a predicted target y1 and an observable hidden variable OH;
m is used to represent the sequence length of the input sequence or the output sequence of the Encoder model, and the length of the input or output sequence in this embodiment is the number of steps of the time step of the input or output of the Encoder model.
And step S13, setting a loss function of the Encoder model. Preferably, the loss function of the Encoder model is set to operate on the weighted average of the loss of the output sequence of the Encoder model.
Step S14, constructing training data of the Encoder model according to the sequence structure set in the step S12, and then training to form the Encoder model by taking the constructed training data as a model training sample; the training data comprises historical time sequence information of the heating and ventilating system and characteristic information of the heating and ventilating system, wherein the characteristic information at least comprises an observable hidden variable OH.
Step S15, cutting an input sequence of the Encode model into (X + OH, M), and cutting an output sequence of the Encode model into (y1+ OH, 1); the input and output sequences of the Encode model are tailored to prepare data for the Decode model prediction. Since the hidden variables required for the prediction of the Decoder model are output step by step in time step, step by step, the input of the Decoder model only requires the output of the last step of the Encoder model, so the number "1" in the output sequence (y1+ OH, 1) of the Encoder model represents the single time step at which the Encoder model is finally output within a certain period of time.
Step S16, constructing a multilayer Decoder model by using an OH-LSTM Cell neural network structure, wherein the construction process is consistent with the Encode model construction process and is not repeated herein;
step S17, setting the input sequence of the Decode model as (X + OH, M) and the output sequence of the Decode model as (y2+ OH, N);
n is used for representing the step number of the time step output by the Decoder model;
y2 is used to represent the prediction target of the Decoder model;
y2+ OH represents that the output of the Decoder model contains a predicted target y2 and an observable hidden variable OH;
step S18, setting a loss function of the Decoder model; preferably, the loss function of the Decoder model is set to average the output sequence loss of the Decoder model.
Step S19, constructing training data of an Encoder-Decoder model according to the sequence structure set in the step S12 and the step S17; the training data comprises historical time sequence information of the heating and ventilating system and characteristic information of the heating and ventilating system, wherein the characteristic information at least comprises an observable hidden variable OH.
Step S20, migrating the model parameters of the Encoder model to the Decode model, and training to form the Encoder-Decode model as a time sequence model by taking the training data constructed in the step S19 as a training sample;
step S21, cutting the input sequence of the Encoder-Decoder model as (X + OH, M + N), cutting the output sequence of the Encoder-Decoder model as (y2, N),
m + N is used to represent that the input sequence of the Encoder-Decoder model contains M + N time steps.
In order to test the performance of the time sequence model, the invention provides a continuous sensitivity curve analysis and evaluation method, please refer to fig. 3, which specifically includes the following steps:
step L1, randomly extracting a sample list of specified continuous step lengths as input of an Encoder model, wherein data in the sample list are training data for constructing a time sequence model; it is also preferable that the number of steps of the consecutive steps is 30, that is, sample data of the heating and ventilation system at consecutive 30 time steps is randomly extracted.
And L2, selecting n sensitivity parameters to be analyzed and evaluated, wherein the sensitivity parameters refer to adjustable control parameters in an industrial scene, such as the frequency of various pumps in the heating and ventilation system, the frequency of a fan and the like.
Step L3, constructing a parameter integer list of the sensitivity parameters; the integer list of sensitivity parameters is made according to the actual control range of the sensitivity parameters, such as the controllable frequency of the fan is [30,50 ].
Step L4, performing Cartesian product combination on the parameter integer list and the input sequence of the Encoder model inputted with the t-th time step in the continuous step specified in the step L1 to construct the input sequence of n groups of Decoder models; if the step size of the specified continuous time step is 30, the value of t is preferably 30, that is, the parameter integer list and the input sequence of the last time step of the specified continuous step size input into the Encoder model are combined by Cartesian product.
L5, carrying out Cartesian product combination on the input sequence of the Encoder model and the input sequence of the n groups of Decoder models to construct the input sequence of the n groups of Encoder-Decoder models; it should be noted here that, instead of directly combining the input sequence of the Encoder model and the input sequence of the n sets of Decoder models, the input sequence of the n sets of Encoder-Decoder models is obtained by performing cartesian product combination on the output sequence of the Encoder model output at the last time step within a specified continuous step and the input sequence of the n sets of Decoder models. Since the output sequence of the Encoder model is predicted from the input sequence, the process of constructing the input sequence of the Encoder-Decoder model is summarized as the cartesian product combination of the input sequence of the Encoder model and the input sequence of the n-group Decoder model for the sake of clarity in step L5. In addition, the reason for constructing n groups of inputs of the Encoder-Decoder model instead of constructing one group of inputs is to increase the confidence of the model performance evaluation result.
And L6, taking the input sequence of the n groups of Encoder-Decoder models constructed in the step L5 as the input of the Encoder-Decoder models, and acquiring the output of the Encoder-Decoder models.
And L7, drawing a time sequence distribution curve which can represent the consumption condition of the heating and ventilation system in a specified time period with continuous step length according to the output of the Encoder-Decoder model.
Please refer to fig. 8 and 9 for a schematic diagram of a time sequence distribution curve obtained by a continuous sensitivity curve analysis and evaluation method.
The present invention further provides a method for evaluating model performance, which is a method for analyzing and evaluating a delay sensitivity curve, please refer to fig. 4, wherein the specific process of evaluating the performance of the time sequence model by the method for analyzing and evaluating the delay sensitivity curve comprises the following steps:
step M1, randomly extracting M sections of sample list groups with appointed continuous step length as the input of the Encoder model; the sample list group comprises m groups of sample lists, and data in the sample lists are training data used for constructing a time sequence model; preferably, the time step (step size) of the specified consecutive step sizes is 30, so that the sufficiency of data can be ensured.
Step M2, n sensitivity parameters to be analyzed and evaluated are selected.
Step M3, construct a parameter integer list of sensitivity parameters.
Step M4, carrying out Cartesian product combination on the parameter integer list and the M groups of input sequences of the Encoder model, and constructing the input sequences of the n x M groups of Decoder models; the construction process of the input sequence of the Decoder model and the step L4 in the continuous sensitivity curve analysis and evaluation method are not described herein again.
Step M5, inputting the sensitivity parameters of each group in the n x M groups and the t-th time step in the specified continuous step length into the input sequence of the Encoder model to carry out Cartesian product combination so as to construct the input sequence of the n x M groups Encoder-Decoder model; t is also preferably 30, i.e. the sensitivity parameters of each of the n x m groups are combined by cartesian products with the input sequence of the last time step of the successive steps of the current time segment to the Encoder model.
And step M6, taking the n x M group input sequence constructed in the step M5 as the input of the Encoder-Decoder model, and acquiring the output of the Encoder-Decoder model.
And step M7, drawing a time sequence distribution curve which can represent the energy consumption condition of the heating and ventilation system in different time periods according to the output of the Encoder-Decoder model.
Fig. 10 is a schematic diagram of a time sequence distribution curve obtained by a delay sensitivity curve analysis and evaluation method.
Referring to fig. 5, in step S3, the process of performing constrained global continuous optimization on the operation strategy of the heating and ventilation system according to the trained timing model specifically includes the following steps:
step N1, extracting historical sample data of the heating and ventilation system at the current moment to construct a control parameter combination candidate set of the Decoder model in j continuous historical time steps, wherein j is a natural number more than or equal to 1; the historical sample data comprises various control parameters of the heating and ventilation system at the historical time step; the current moment refers to the moment when the heating and ventilation system needs to be globally and continuously optimized;
step N2, constructing k groups of control parameter combination candidate sets; a group of control parameter combination candidate sets comprises control parameters of specified time steps, such as control parameters of j time steps, and the control parameters of each time step may not be consistent;
step N3, combining k groups of control parameter combination candidate sets with environmental parameters to construct an input sequence of m groups of Decoder models; the environmental parameters are temperature and humidity information of the operation of the heating and ventilation system and the like; the combination method of the control parameter combination candidate set and the environmental parameters is to carry out mathematical splicing on the two, and the existing data splicing modes are many, so the specific splicing process is not explained here;
step N4, combining the constructed input sequence of the k group Decoder model with the input sequence of the Encoder model which is also k group, and constructing the input sequence of the k group Encoder-Decoder model; as for the method for constructing the input sequence of the Encoder-Decoder model, such as the method for analyzing and evaluating the continuous sensitivity curve or the method for analyzing and evaluating the time delay sensitivity curve, is not described herein again;
step N5, inputting the constructed input sequence of the k groups of Encoder-Decoder models into the Encoder-Decoder models, and outputting the prediction result of the output power of the heating and ventilation system;
step N6, removing the control parameter combination candidate set which does not meet the constraint condition according to the prediction result; the elimination rule is that each control parameter has a reasonable control range, and when the prediction result shows that the parameter value corresponding to the control parameter exceeds the reasonable control range, the control parameter combination candidate set where the control parameter is located is eliminated;
step N7, selecting the control parameter combination candidate set with the lowest predicted output power of the heating and ventilation system from the at least one group of control parameter combination candidate sets which are in accordance with the constraint conditions as the control parameter set of the heating and ventilation system at the next time step at the current moment, and issuing the control parameter combination candidate set to corresponding equipment in the heating and ventilation system, wherein each equipment in the heating and ventilation system operates according to the issued control parameters;
and step N8, repeatedly executing N1-N7 in the next time step of the current time so as to make continuous optimization decision on the global operation state of the heating and ventilation system.
It should be understood that the above-described embodiments are merely preferred embodiments of the invention and the technical principles applied thereto. It will be understood by those skilled in the art that various modifications, equivalents, changes, and the like can be made to the present invention. However, such variations are within the scope of the invention as long as they do not depart from the spirit of the invention. In addition, certain terms used in the specification and claims of the present application are not limiting, but are used merely for convenience of description.

Claims (10)

1. A method for continuously deciding energy efficiency optimization of a data center based on seq2seq is characterized by comprising the following steps:
step S1, establishing a time sequence model of the water circulation of the heating and ventilation system based on the seq2seq structure of the recurrent neural network;
step S2, evaluating the performance of the time sequence model;
and step S3, carrying out constrained continuous operation strategy optimization on the whole situation of the heating and ventilation system by utilizing the time sequence model with the performance meeting the expectation.
2. The energy efficiency optimization continuous decision method according to claim 1, wherein in the step S1, the constructing the time series model specifically includes the following steps:
step S11, building a multilayer Encoder model by using an OH-LSTM Cell neural network structure;
step S12, setting the input sequence of the Encoder model as (X + OH, M) and the output sequence of the Encoder model as (y1+ OH, M);
x is used for representing historical time sequence information of the heating and ventilating system in a specified time period;
OH is used for representing an observable hidden variable of the heating and ventilation system in the specified time period;
x + OH is the combined characteristic of the historical time sequence information and characteristic information combination input into the Encoder model in the specified time period;
y1 is used to represent the prediction target of the Encoder model;
y1+ OH represents that the output of the Encoder model includes the predicted target y1 and the observable hidden variable OH;
m is used for representing the step number of the time step of the input sequence or the output sequence of the Encoder model;
step S13, setting a loss function of the Encoder model;
step S14, constructing the training data of the Encoder model according to the sequence structure set in the step S12, and then training to form the Encoder model by taking the training data as a model training sample;
s15, cutting an input sequence of the Encoder model into (X + OH, M), and cutting an output sequence of the Encoder model into (y1+ OH, 1);
"1" represents a single said time step that the Encoder model last outputs;
step S16, constructing a multilayer Decoder model by using the OH-LSTM Cell neural network structure;
step S17, setting the input sequence of the Decode model as (X + OH, M) and the output sequence of the Decode model as (y2+ OH, N);
n is used for representing the step number of the time step output by the Decoder model;
y2 is used to represent the prediction target of the Decoder model;
y2+ OH denotes that the output of the Decoder model includes the predicted target y2 and the observable hidden variable OH;
step S18, setting a loss function of the Decoder model;
step S19, constructing training data of an Encoder-Decoder model according to the sequence structure set in the step S12 and the step S17;
step S20, transferring the model parameters of the Encoder model to the Decode model, and training to form the Encoder-Decode model as the timing sequence model by taking the training data constructed in the step S19 as a training sample;
step S21, cutting the input sequence of the Encoder-Decoder model into (X + OH, M + N), cutting the output sequence of the Encoder-Decoder model into (y2, N),
m + N is used to represent that the input sequence of the Encoder-Decoder model contains M + N time steps.
3. The energy efficiency optimization continuous decision method according to claim 2, characterized in that the characteristic information comprises at least the observable hidden variables.
4. The energy efficiency optimization continuous decision method according to claim 2, wherein in the step S13, the loss function of the Encoder model is set to perform a weighted average operation on the loss of the output sequence of the Encoder model.
5. The energy efficiency optimization continuous decision method according to claim 2, wherein in the step S18, the loss function of the Decoder model is set to average the output sequence loss of the Decoder model.
6. The energy efficiency optimization continuous decision method according to claim 2, wherein in step S2, the method for performing performance evaluation on the time series model includes a continuous sensitivity curve analysis and evaluation method, and the continuous sensitivity curve analysis and evaluation method specifically includes the following steps:
step L1, randomly extracting a sample list of specified continuous step sizes as the input of the Encoder model;
step L2, selecting n sensitivity parameters to be analyzed and evaluated;
step L3, constructing a parameter integer list of the sensitivity parameters;
a step L4 of cartesian product combining the parameter integer list with the input sequence of the Encoder model to which the t-th time step within the consecutive step sizes specified by the step L1 is input, to construct n sets of input sequences of the Decoder model;
l5, carrying out Cartesian product combination on the input sequence of the Encoder model and the input sequences of n groups of the Decoder model, and constructing n groups of input sequences of the Encoder-Decoder model;
a step L6 of taking the input sequence of the n sets of Encoder-Decoder models constructed in the step L5 as the input of the Encoder-Decoder models and obtaining the output of the Encoder-Decoder models;
and L7, drawing a time sequence distribution curve which can represent the energy consumption situation of the heating and ventilation system in the specified time period of the continuous step length according to the output of the Encoder-Decoder model.
7. The energy efficiency optimization continuous decision method according to claim 2, wherein in step S2, the method for performing performance evaluation on the time series model includes a delay sensitivity curve analysis and evaluation method, and the delay sensitivity curve analysis and evaluation method specifically includes the following steps:
step M1, randomly extracting M sections of sample list groups with appointed continuous step length as the input of the Encoder model;
step M2, selecting n sensitivity parameters to be analyzed and evaluated;
step M3, constructing a parameter integer list of the sensitivity parameters;
step M4, carrying out Cartesian product combination on the parameter integer list and M groups of input sequences of the Encoder model to construct n x M groups of input sequences of the Decoder model;
a step M5 of inputting the sensitivity parameters of each group in the n x M groups and the t-th time step in the specified continuous step into the input sequence of the Encoder model to carry out Cartesian product combination so as to construct n x M groups of input sequences of the Encoder-Decoder model;
a step M6 of taking the input sequence of the Encoder-Decoder model of the n x M groups constructed in the step M5 as the input of the Encoder-Decoder model and obtaining the output of the Encoder-Decoder model;
and step M7, drawing a time sequence distribution curve which can represent the energy consumption condition of the heating and ventilation system in different time periods according to the output of the Encoder-Decoder model.
8. The energy efficiency optimization continuous decision method according to claim 6 or 7, wherein the number of steps of the continuous step size randomly extracted is 30 time steps.
9. The energy efficiency optimization continuous decision method according to claim 6 or 7, characterized in that t is 30.
10. The energy efficiency optimization continuous decision method according to claim 2, wherein the step S3, the process of performing constrained global continuous optimization on the operation strategy of the heating and ventilation system according to the time sequence model comprises the following steps:
step N1, extracting historical sample data of the heating and ventilation system at the current moment to construct a control parameter combination candidate set of the Decoder model in j continuous historical time steps, wherein j is a natural number more than or equal to 1;
step N2, constructing k groups of control parameter combination candidate sets;
step N3, combining k groups of control parameter combination candidate sets with environmental parameters to construct m groups of input sequences of the Decoder model;
step N4, combining the constructed input sequences of k groups of the Decoder model with the input sequences of the Encoder model which is also k groups, and constructing the input sequences of k groups of the Encoder-Decoder model;
a step N5 of inputting the input sequence of the k sets of the Encoder-Decoder models constructed in the step N4 into the Encoder-Decoder models and outputting a prediction result of the output power of the heating and ventilation system;
step N6, according to the prediction result, eliminating the control parameter combination candidate set which does not meet the constraint condition;
step N7, selecting the control parameter combination candidate set with the lowest predicted output power of the heating and ventilation system from at least one group of the control parameter combination candidate sets which are in accordance with the constraint condition as the control parameter set of the heating and ventilation system at the next time step of the current time, and issuing the control parameter set to corresponding equipment in the heating and ventilation system, wherein each equipment in the heating and ventilation system operates according to the issued control parameter;
and step N8, repeating the steps N1-N7 in the next time step of the current time so as to make continuous optimization decision on the global operation state of the heating and ventilation system.
CN202010122279.6A 2020-02-27 2020-02-27 Data center energy efficiency optimization continuous decision method based on seq2seq Active CN111368408B (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202310735629.XA CN116611347A (en) 2020-02-27 2020-02-27 Continuous decision method for energy efficiency optimization of data center
CN202010122279.6A CN111368408B (en) 2020-02-27 2020-02-27 Data center energy efficiency optimization continuous decision method based on seq2seq

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010122279.6A CN111368408B (en) 2020-02-27 2020-02-27 Data center energy efficiency optimization continuous decision method based on seq2seq

Related Child Applications (1)

Application Number Title Priority Date Filing Date
CN202310735629.XA Division CN116611347A (en) 2020-02-27 2020-02-27 Continuous decision method for energy efficiency optimization of data center

Publications (2)

Publication Number Publication Date
CN111368408A true CN111368408A (en) 2020-07-03
CN111368408B CN111368408B (en) 2023-06-23

Family

ID=71210048

Family Applications (2)

Application Number Title Priority Date Filing Date
CN202010122279.6A Active CN111368408B (en) 2020-02-27 2020-02-27 Data center energy efficiency optimization continuous decision method based on seq2seq
CN202310735629.XA Pending CN116611347A (en) 2020-02-27 2020-02-27 Continuous decision method for energy efficiency optimization of data center

Family Applications After (1)

Application Number Title Priority Date Filing Date
CN202310735629.XA Pending CN116611347A (en) 2020-02-27 2020-02-27 Continuous decision method for energy efficiency optimization of data center

Country Status (1)

Country Link
CN (2) CN111368408B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117404853A (en) * 2023-12-14 2024-01-16 山西省水利建筑工程局集团有限公司 External circulating water cooling system and method for tunnel boring machine

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109284576A (en) * 2018-10-29 2019-01-29 东北电力大学 A kind of distributing electric heating load scheduling method and its modeling based on measured data
CN110703899A (en) * 2019-09-09 2020-01-17 创新奇智(南京)科技有限公司 Data center energy efficiency optimization method based on transfer learning

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109284576A (en) * 2018-10-29 2019-01-29 东北电力大学 A kind of distributing electric heating load scheduling method and its modeling based on measured data
CN110703899A (en) * 2019-09-09 2020-01-17 创新奇智(南京)科技有限公司 Data center energy efficiency optimization method based on transfer learning

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
黄莉等: "一种面向传统水冷式中央空调主机系统能效优化控制方法" *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117404853A (en) * 2023-12-14 2024-01-16 山西省水利建筑工程局集团有限公司 External circulating water cooling system and method for tunnel boring machine
CN117404853B (en) * 2023-12-14 2024-03-08 山西省水利建筑工程局集团有限公司 External circulating water cooling system and method for tunnel boring machine

Also Published As

Publication number Publication date
CN111368408B (en) 2023-06-23
CN116611347A (en) 2023-08-18

Similar Documents

Publication Publication Date Title
CN113835344B (en) Control optimization method of equipment, display platform, cloud server and storage medium
CN107609667B (en) Heat supply load prediction method and system based on Box _ cox transformation and UFCNN
Deng et al. Integrating feature optimization using a dynamic convolutional neural network for chemical process supervised fault classification
CN112415924A (en) Energy-saving optimization method and system for air conditioning system
CN113205203A (en) CNN-LSTM-based building energy consumption prediction method and system
CN115370437A (en) Cold, heat and electricity tri-generation energy storage system based on compressed fluid and control method
CN111368408A (en) Seq2 seq-based data center energy efficiency optimization continuous decision method
CN110033126A (en) Shot and long term memory network prediction technique based on attention mechanism and logistic regression
CN116578858A (en) Air compressor fault prediction and health degree evaluation method and system based on graphic neural network
CN117113054A (en) Multi-element time sequence prediction method based on graph neural network and transducer
CN115828769A (en) Method for predicting working condition of cooling tower and reducing consumption based on intelligent calculation
CN116108749A (en) Wind turbine flow field prediction method based on deep learning
Wang Application of deep learning model in building energy consumption prediction
CN116562111A (en) Data center energy saving method, device, system and storage medium
CN114169605A (en) Physical field and performance collaborative prediction method suitable for pump and hydraulic turbine
Zhu et al. Application of machine learning and its improvement technology in modeling of total energy consumption of air conditioning water system
CN117634217B (en) Intelligent building management system and method based on digital twinning
CN114650190B (en) Energy-saving method, system, terminal equipment and storage medium for data center network
Jradi et al. A Digital Twin platform for energy efficient and smart buildings applications
Qin et al. Cnn-gru-attention based short-term load forecasting of distribution networks
HUANG et al. Energy-driven intelligent generative urban design
CN117370870B (en) Knowledge and data compound driven equipment multi-working condition identification and performance prediction method
CN111523641B (en) Sector delay prediction method based on ConvLSTM-SRU
Chen et al. Integrated attention mechanism for GBDT building energy consumption prediction algorithm
Ma et al. Monotonic neural network: combining deep learning with domain knowledge for chiller plants energy optimization

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