CN111637614A - Intelligent control method for data center active ventilation floor - Google Patents

Intelligent control method for data center active ventilation floor Download PDF

Info

Publication number
CN111637614A
CN111637614A CN202010455152.6A CN202010455152A CN111637614A CN 111637614 A CN111637614 A CN 111637614A CN 202010455152 A CN202010455152 A CN 202010455152A CN 111637614 A CN111637614 A CN 111637614A
Authority
CN
China
Prior art keywords
time
phi
rack
eval
active ventilation
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
CN202010455152.6A
Other languages
Chinese (zh)
Other versions
CN111637614B (en
Inventor
万剑雄
周杰
熊伟
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.)
Inner Mongolia University of Technology
Original Assignee
Inner Mongolia University of Technology
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 Inner Mongolia University of Technology filed Critical Inner Mongolia University of Technology
Priority to CN202010455152.6A priority Critical patent/CN111637614B/en
Publication of CN111637614A publication Critical patent/CN111637614A/en
Application granted granted Critical
Publication of CN111637614B publication Critical patent/CN111637614B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F24—HEATING; RANGES; VENTILATING
    • F24F—AIR-CONDITIONING; AIR-HUMIDIFICATION; VENTILATION; USE OF AIR CURRENTS FOR SCREENING
    • F24F11/00—Control or safety arrangements
    • F24F11/89—Arrangement or mounting of control or safety devices
    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F24—HEATING; RANGES; VENTILATING
    • F24F—AIR-CONDITIONING; AIR-HUMIDIFICATION; VENTILATION; USE OF AIR CURRENTS FOR SCREENING
    • F24F11/00—Control or safety arrangements
    • F24F11/62—Control or safety arrangements characterised by the type of control or by internal processing, e.g. using fuzzy logic, adaptive control or estimation of values
    • F24F11/63—Electronic processing
    • F24F11/64—Electronic processing using pre-stored data

Landscapes

  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Chemical & Material Sciences (AREA)
  • Combustion & Propulsion (AREA)
  • Mechanical Engineering (AREA)
  • General Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Air Conditioning Control Device (AREA)

Abstract

An intelligent control method for an active ventilation floor of a data center establishes a Markov decision process model for a hot spot problem of a rack of the data center and provides three model solving algorithms, including a basic intelligent algorithm, a sample value variant intelligent algorithm and a structure variant intelligent algorithm, which are respectively used as the cores of the active ventilation floor control algorithm. The model consists of four parts, namely a system state, a behavior, an incentive and a value function, the solution of the model is that the optimal behavior is continuously selected under a series of system states, so that the accumulated incentive of the system is maximized, the active ventilation floor control algorithm is adopted, the complex relation between the temperature distribution of the air inlet of the rack and the rotating speed of the active ventilation floor fan is continuously explored and learned, the optimal PWM signal duty ratio value can be finally generated according to the temperature distribution of the air inlet of the rack, the rotating speed of the active ventilation floor fan is adjusted, the temperature distribution of the air inlet of the rack is uniform, and the hot spot problem of the rack is relieved. Compared with other schemes, the method has higher universality, is easier to deploy and has more cost benefit.

Description

Intelligent control method for data center active ventilation floor
Technical Field
The invention belongs to the technical field of automatic control, and particularly relates to an intelligent control method for an active ventilation floor of a data center.
Background
The rack hot spot is a high-temperature point at which the temperature of one or more positions of the rack of the data center machine room is obviously higher than that of other positions. Excessive temperatures can cause some servers in a data center to operate less efficiently, thereby reducing its overall power density and also reducing its reliability, which is clearly contrary to the needs of data centers.
The hot spots of the racks are relieved or eliminated by adopting a global regulation and control mode, for example, the power of an air conditioner in a machine room is increased to provide sufficient cold air, so that most of the rack areas are in an over-cooling state inevitably, and the total energy consumption of the data center is more huge than half of the total energy consumption of the data center while the waste of cooling resources is caused. Thus, rack level cooling solutions are more suitable for mitigating rack hot spot issues.
There are currently rack-level refrigeration solutions, such as installing adaptive ventilation floors, installing baffles, enclosing individual racks and providing them with ventilation ducts, etc. However, these solutions are "passive" cooling solutions, which do not actively provide a cooling air flow to the racks, and they are not sufficient when the cooling air supply is insufficient.
The active ventilation floor is used as another rack-level refrigeration scheme, the hot spot problem of the rack is relieved by actively conveying cold air, and compared with the scheme, the active ventilation floor is easier to deploy and more cost-effective, but the control difficulty is mainly characterized by the diversity and the dynamic property of the placing environment, such as different distribution of machine room air conditioners, relative positions of racks and servers in the racks; the cold and hot channels are in different closed states, and the server rack is in different standards and sealing conditions; the room air conditioning power, the thermal load of different rack servers, etc. Therefore, the thermal energy efficiency and airflow model of the data center is generally difficult to describe by an analytic model.
Most of the existing active ventilation floor related researches are performance modeling and evaluation based on measurement or simulation, and no research literature of active ventilation floor control problems exists at present.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide an intelligent control method for an active ventilation floor of a data center, which automatically learns an optimal operation strategy and plans the air flow of a rack on the premise of not increasing the power of an air conditioner of a machine room, so that the temperature distribution of an air inlet of the rack is uniform, and the hot spot problem of the rack is relieved. And complex airflow and heat exchange models do not need to be established and calibrated, so that the universality of the active ventilation floor is improved.
In order to achieve the purpose, the invention adopts the technical scheme that:
an intelligent control method for an active ventilation floor of a data center establishes a Markov decision process model for a hot spot problem of a rack of the data center and provides three model solving algorithms, including a basic intelligent algorithm, a sample value variant intelligent algorithm and a structure variant intelligent algorithm, which are respectively used as the cores of the active ventilation floor control algorithm. The model consists of four parts, namely a system state, a behavior, an incentive and a value function, the solution of the model is that the optimal behavior is continuously selected under a series of system states, so that the accumulated incentive of the system is maximized, the active ventilation floor control algorithm can generate the optimal PWM signal duty ratio value according to the temperature distribution of the air inlet of the rack and adjust the rotating speed of the active ventilation floor fan by continuously exploring and learning the complex relation between the temperature distribution of the air inlet of the rack and the rotating speed of the active ventilation floor fan, so that the temperature distribution of the air inlet of the rack is uniform, and the hot spot problem of the rack is relieved.
Compared with the prior art, the invention has the beneficial effects that:
the invention does not need to establish and calibrate complex airflow and heat exchange models, uses an intelligent control algorithm, overcomes the diversity and the dynamic property of the placing environment of the active ventilation floor, automatically matches the temperature distribution of the air inlet of the rack and the rotating speed of the fan of the optimal active ventilation floor, only needs to replace the original common ventilation floor with the active ventilation floor for operating the invention, can operate autonomously, improves the temperature distribution of the air inlet of the rack, and relieves the hot spot problem of the rack.
Drawings
FIG. 1 is a diagram of active vent floor design and deployment.
Detailed Description
The embodiments of the present invention will be described in detail below with reference to the drawings and examples.
Fig. 1 is a detailed deployment implementation schematic diagram of the invention, wherein a certain number of temperature sensors 1 are uniformly distributed at an air inlet of a rack 2 to monitor the temperature distribution of the air inlet of the rack 2, and a temperature sensor two is additionally arranged below an active ventilation floor to monitor the air supply temperature below the active ventilation floor.
In the field, the rack 2 is a rectangular iron box, a certain number of servers are placed in the box, and a plurality of racks are placed in a row. In a certain row of racks, a left panel and a right panel of a certain rack are generally tightly attached to other racks, a front panel of the rack is an air inlet for sucking cold air to refrigerate the server, a rear panel of the rack is an air outlet for discharging refrigerated hot air, the temperature distribution of the air inlet of the rack is monitored, namely the temperature of certain positions of the front panel of the rack is monitored, the temperature of the positions forms the temperature distribution of the air inlet of the rack, and therefore the number of the first temperature sensors 1 depends on the number of the positions.
The intelligent control method of the active ventilation floor is operated at a PC end, the PC6 is connected with the microcontroller 3, the microcontroller 3 is connected with the drive board 4, and the drive board 4 is connected with the active ventilation floor fan 7 after being connected with the switching power supply 5(12V, 20A). According to the temperature distribution returned by the temperature sensor I1, a duty ratio value of a PWM signal is generated and transmitted to the microcontroller 3, the microcontroller 3 generates a corresponding PWM signal according to the duty ratio value and transmits the PWM signal to the drive plate 4, the drive plate 4 controls the voltage provided by the switching power supply 5 to the active ventilation floor fan 7 according to the PWM signal, and the purpose of adjusting the rotating speed of the fan is achieved by controlling the power supply voltage of the fan.
The control method comprises the following steps:
1. a Markov decision process model is established for the hot spot problem of a rack of a data center of a raised floor structure (a data center air supply structure, a data center machine room floor is elevated, and a 60-100cm high underfloor space is reserved for conveying cold air by a machine room air conditioner, the structure is the raised floor structure, and most of domestic data centers adopt the structure at present), and the raised floor structure consists of the following four parts:
the system A state is a rack air inlet temperature distribution set with history, and the formula is as follows:
φt＝{st-p,…,sx…,st-1,sttherein of
Figure BDA0002508955000000031
Wherein phitSystem state at time t, st-p、sx、st-1、stThe temperature distribution of the air inlet of the frame at t-p, x, t-1 and t moments, x ∈ [ t-p, t [ ]]P is the history length; t isiIs the reading of the temperature sensor one numbered i,
Figure BDA0002508955000000032
Figure BDA0002508955000000033
is a collection of the first temperature sensors,
Figure BDA0002508955000000034
is the total number of the temperature sensors one.
Space of B behavior
Figure BDA0002508955000000041
Defined as the duty ratio value of the discretized PWM signal, the formula is:
Figure BDA0002508955000000042
wherein a is
Figure BDA0002508955000000043
In a certain action, DC is the duty ratio of PWM signal, max (DC) is the maximum duty ratio, DDRLFor DC discretization of equivalence ratios, k denotes D in a certain behaviorDRLThe number of (2);
c award Rt+1The temperature distribution uniformity of the air inlet of the rack is quantified, and the energy consumption of the active ventilation floor fan is calculated according to the following formula:
Figure BDA0002508955000000044
wherein R ist+1The reward obtained after the system takes some action for time t,
Figure BDA0002508955000000045
the temperature distribution uniformity of the air inlet of the rack is shown, the value of the formula is all negative, the closer to 0, the more uniform the temperature distribution of the air inlet of the rack is shown, wherein Tt,iThe temperature reading of sensor one numbered i at time t,
Figure BDA0002508955000000046
for the reference temperature of the rack at time t,
Figure BDA0002508955000000047
Tt,underis the reading of the second temperature sensor at time t, ΔTThe fixed temperature difference is set according to the mixing degree of the cold air flow and the hot air flow on the active ventilation floor, and is positive; - (A)ref×DCt)3Representing the active ventilation floor fan energy consumption, the values of the formula are all negative, the closer to 0, the lower the fan energy consumption, wherein ArefTo maintain a reference behavior value of the same order of magnitude as the uniformity of the temperature distribution at the inlet of the frame, DCtThe square wave duty ratio of the PWM signal at time t.
D value function Q (phi)t,at) The formula of the behavior cost function is as follows:
Figure BDA0002508955000000048
wherein the cost function Q (phi)t,at) Referred to as the Q-function,
Figure BDA0002508955000000049
for the action taken by the system at time t,
Figure BDA00025089550000000410
as a function of the expectation, y is the future time relative to time t, Rt+y+1Represents the reward obtained after the system takes action at the time of t + y, gamma represents the attenuation factor and represents the attention degree of taking certain action under a certain state to the future reward of the system, namely the environmental influence, gamma is more than or equal to 0 and less than 1, and gammayY power of gamma, is t + y time Rt+y+1The attenuation factor of (2).
The E-model can be summarized as maximizing the cumulative reward by selecting the optimal behavior at any time t system state, with the model formula:
Figure BDA0002508955000000051
is constrained to
Figure BDA0002508955000000052
Wherein, γtIs time t system Rt+1The attenuation factor of (2).
2. Model solution and solving algorithm
a, calculating to obtain an optimal Q function, namely selecting an optimal behavior according to the optimal Q function under the system state at any time t to maximize the accumulated reward, wherein the optimal Q function calculation formula is as follows:
Figure BDA0002508955000000053
at any time t, the optimal behavior selection formula is as follows:
Figure BDA0002508955000000054
wherein Q*(φt,at) Represents the optimal Q function, phit+1Represents the state of the system at time t +1, and a represents any of all possible actions taken by the system at time t +1, i.e., the action space
Figure BDA0002508955000000055
A certain behavior in (2).
And b, solving an algorithm, namely, calculating to obtain an optimal Q function and selecting an optimal behavior in the decision so as to maximize the accumulated reward. The solving algorithm comprises a basic intelligent algorithm, a sample value variant intelligent algorithm and a structure variant intelligent algorithm, wherein the three algorithms are accumulated through continuous decision (phi)t,at,Rt+1,φt+1) Training the neural network with the sample records such that the neural network approximates a Q function, thereby selecting an optimal behavior that maximizes the cumulative reward for the model, where φtIs the system state at time t, atFor the action taken by the system at time t, Rt+1Adopt a for the systemtThe reward obtained after, phit+1The system state at time t + 1. The three algorithms are designed as follows:
a) the basic intelligent algorithm is characterized in that two neural networks with the same structure are used for approximating a Q function, one is used for approximating the Q sample function, and a Q sample value is calculated and is called as a targ network; the other is used for approximating a Q prediction function and calculating a Q prediction value, and is called eval network; and calculating the difference between the Q sample value and the Q predicted value by using the sample record, and training and updating the neural network, wherein the Q sample value calculation formula is as follows:
Figure BDA0002508955000000061
wherein Qt+1,targetIs a value of Q sample, Rt+1And phit+1Gamma of 0 or more and less than 1 is an attenuation factor, Q (phi), taken from the sample recordt+1,a；θt,target) The set of Q samples output by targ network, a represents all possible actions the system might take at time t +1,
Figure BDA0002508955000000062
as a space of action, θt,targetIs a set of network parameters at time targ.
The neural network updating mode is as follows:
Figure BDA0002508955000000063
whereint+1Is the difference between the Q sample value and the corresponding Q predicted value, Q (phi)t,at；θt,eval) In the set of Q predictions output for eval network, atCorresponding predicted value of Q, phitAnd atTaken from the sample record, θt,evalFor the eval network time t parameter set, thetat+1,evalFor the eval network time t +1 parameter set,
Figure BDA0002508955000000064
is composed of
Figure BDA0002508955000000065
About thetat,evalα is the neural network learning step size, θtargetWhen the time t is an integral multiple (including 0) of N, targ network parameter set thetaevalWhen the time t is an integer multiple (including 0) of N, the eval network parameter set.
b) The sample value variant intelligent algorithm uses the formula in calculating the Q sample value:
Figure BDA0002508955000000066
wherein Qt+1,targetIs a value of Q sample, Rt+1And phit+1Taken from the sample record, Q (phi)t+1,a；θt,target) For the set of Q samples output by the targ network,
Figure BDA0002508955000000071
for the Q sample set output by targ network, let Qeval(φt+1,a；θt,eval) Q sample value, Q, corresponding to the largest behavioreval(φt+1,a；θt,eval) For the Q prediction set output by the eval network, a represents any one of all possible actions taken by the system at the moment t +1, namely, the action space
Figure BDA0002508955000000072
A certain behavior of, thetat,evalSet of eval network parameters for time tAnd thetat,targetA targ network parameter set at the time t;
the neural network structure and the updating mode of the sample value variant intelligent algorithm are the same as those of the basic intelligent algorithm.
c) The structure variant intelligent algorithm uses two neural networks with the same structure, a DN layer is arranged on the penultimate layer of each neural network, the DN layer is divided into a V section and an A section, the node number of the neurons in the V section is 1 and represents the system state at the moment t, the number of the neurons in the A section is the number of elements in the action space and represents all actions which can be taken under the system state, and the calculation formula of the DN layer is as follows:
Figure BDA0002508955000000073
wherein, Q (phi)t,at；θt,θt,V,θt,A) Is the final output of the neural network, phitAnd atTaken from the sample record, θtFor time t, a network parameter set theta before DN layer of the neural network of the structure variant intelligent algorithmt,VFor the DN layer V-segment parameter at time t, θt,AFor the DN segment A parameter at time t, V (phi)t；θt,θt,V) For the output of V section, A (phi)t,at；θt,θt,A) Is a in section AtCorresponding output value, A (phi)t,a'；θt,θt,A) For all outputs of segment A, a' represents the state φtIn the following, all actions that the system may take,
Figure BDA0002508955000000074
the number of elements in the action space.
And then, training and updating the neural network by adopting a Q sample value calculation and neural network updating mode which is the same as the sample value variant intelligent algorithm.
3. Through continuously exploring and learning the complex relation between the temperature distribution of the air inlet of the rack and the rotating speed of the active ventilation floor fan, the optimal duty ratio value of a PWM (pulse-width modulation) signal is finally generated according to the temperature distribution of the air inlet of the rack, and the rotating speed of the active ventilation floor fan is adjusted, so that the temperature distribution of the air inlet of the rack is uniform, and the problem of hot spots of the rack is solved. The running logic of the system at the PC end is as follows:
1: in different control algorithms, different neural networks are constructed and initialized, and targ network parameters are the same as eval network parameters; setting the sample record cache array; setting a reference temperature Tt；
2: setting an initial time t to be 0, and recording the time recorded by the sample in the cache array as tau; probability of exploration of initial behavior, rate of exploration decreasing with tMinimum probability of explorationmin；
3: randomly selecting behaviors in Z moments and recording (phi) generated at each momentz∈[0,Z),az∈[0,Z),Rz+1∈[0,Z],φz+1∈[0,Z]) Storing the data into a cache array;
4: obtaining the temperature distribution of the air inlet of the initial frame
Figure BDA0002508955000000081
5: starting a circulation body;
6: obtaining p historical frame air inlet temperature distributions to jointly form a system state phit＝{st-p,…,st-1,st}；
7: if t is equal to 0, then the action a is selectedtMax (dc) and go 9, otherwise go 8;
8: behavior is selected using the following formula:
Figure BDA0002508955000000082
9: execution of atThe PC sends a duty ratio instruction to the microcontroller, changes the rotating speed of the fan and obtains the temperature distribution s of the air inlet of the rack at the next moment of the systemt+1Calculating R according to the winning incentive formula of claim 4t+1；
10: forming a next state phi according to the latest p pieces of temperature distribution historyt+1＝{st+1-p,…,st,st+1And will be(φt,at,Rt+1,φt+1) Storing the data into a cache array;
11: randomly extracting Y sample records (phi) from the buffer arrayτ,aτ,Rτ+1,φτ+1)；
12: and calculating the Q sample value by utilizing Y records according to different control algorithms, wherein the formula is as follows:
Figure BDA0002508955000000083
13: updating the eval network using the learning step size α and the loss function:
Figure BDA0002508955000000084
14: exploration probability of-DeltaAndminminimum value of (1);
15: if t mod N is 0, the targ network copies eval network parameters, otherwise 16 is carried out;
16: the time t is increased by 1;
17: the cycle body is ended.
In conclusion, the invention establishes a Markov decision process model for the hot spot problem of the data center rack, and provides three model solving algorithms, including a basic intelligent algorithm, a sample value variant intelligent algorithm and a structure variant intelligent algorithm, which are respectively used as the core of an active ventilation floor control algorithm. The model consists of four parts, namely a system state, a behavior, an incentive and a value function, the solution of the model is that the optimal behavior is continuously selected under a series of system states, so that the accumulated incentive of the system is maximized, the active ventilation floor control algorithm is adopted, the complex relation between the temperature distribution of the air inlet of the rack and the rotating speed of the active ventilation floor fan is continuously explored and learned, the optimal PWM signal duty ratio value can be finally generated according to the temperature distribution of the air inlet of the rack, the rotating speed of the active ventilation floor fan is adjusted, the temperature distribution of the air inlet of the rack is uniform, and the hot spot problem of the rack is relieved. Compared with other schemes, the method has higher universality, is easier to deploy and has more cost benefit.

Claims (6)

1. The intelligent control method of the data center active ventilation floor is characterized by comprising the following steps:
step 1, arranging a certain number of first temperature sensors for monitoring the temperature distribution of an air inlet of a rack at the air inlet of the rack, and arranging a second temperature sensor for monitoring the air supply temperature under an active ventilation floor under the active ventilation floor;
step 2, establishing a Markov decision process model for the hot spot problem of the data center frame, wherein the model is determined by the system state phitBehavior space
Figure FDA0002508954990000011
Reward Rt+1And a cost function Q (phi)t,at) The four parts are formed;
wherein: system state phi at time ttThe temperature distribution set of the rack air inlet with history is defined, and the formula is as follows:
φt＝{st-p,…,sx…,st-1,sttherein of
Figure FDA0002508954990000012
Wherein s ist-p、sx、st-1、stThe temperature distribution of the air inlet of the frame at t-p, x, t-1 and t moments, x ∈ [ t-p, t [ ]]P is the history length; t isiIs the reading of the temperature sensor one numbered i,
Figure FDA0002508954990000013
Figure FDA0002508954990000014
is a collection of the first temperature sensors,
Figure FDA0002508954990000015
the total number of the temperature sensors is one;
space of action
Figure FDA0002508954990000016
Defined as the duty ratio value of the discretized PWM signal, the formula is:
Figure FDA0002508954990000017
wherein a is
Figure FDA0002508954990000018
In a certain action, DC is the duty ratio of PWM signal, max (DC) is the maximum duty ratio, DDRLFor DC discretization of equivalence ratios, k denotes D in a certain behaviorDRLThe number of (2);
reward Rt+1The temperature distribution uniformity of the air inlet of the rack is quantified, and the energy consumption of the active ventilation floor fan is calculated according to the following formula:
Figure FDA0002508954990000019
wherein R ist+1The reward obtained after the system takes some action for time t,
Figure FDA0002508954990000021
the temperature distribution uniformity of the air inlet of the rack is shown, the value of the formula is all negative, the closer to 0, the more uniform the temperature distribution of the air inlet of the rack is shown, wherein Tt,iThe temperature reading of sensor one numbered i at time t,
Figure FDA0002508954990000022
for the reference temperature of the rack at time t,
Figure FDA0002508954990000023
Tt,underis the reading of the second temperature sensor at time t, ΔTThe fixed temperature difference is set according to the mixing degree of the cold air flow and the hot air flow on the active ventilation floor, and is positive; - (A)ref×DCt)3Representing the active ventilation floor fan energy consumption, the values of the formula are all negative, the closer to 0, the lower the fan energy consumption, wherein ArefTo maintain a reference behavior value of the same order of magnitude as the uniformity of the temperature distribution at the inlet of the frame, DCtThe duty ratio of the square wave of the PWM signal at the moment t;
value function Q (phi)t,at) The formula of the behavior cost function is as follows:
Figure FDA0002508954990000024
wherein the cost function Q (phi)t,at) Referred to as the Q-function,
Figure FDA0002508954990000025
for the action taken by the system at time t,
Figure FDA0002508954990000026
as a function of the expectation, y is the future time relative to time t, Rt+y+1Represents the reward obtained after the system takes action at the time of t + y, gamma represents the attenuation factor and represents the attention degree of taking certain action under a certain state to the future reward of the system, namely the environmental influence, gamma is more than or equal to 0 and less than 1, and gammayY power of gamma, is t + y time Rt+y+1The attenuation factor of (d);
the markov decision process model is summarized as: under the system state at any time t, the system accumulated reward is maximized by selecting the optimal behavior, and the formula is as follows:
Figure FDA0002508954990000027
is constrained to
Figure FDA0002508954990000028
Wherein, γtIs time t system Rt+1The attenuation factor of (d);
and 3, solving the model, continuously exploring and learning the complex relation between the temperature distribution of the air inlet of the rack and the rotating speed of the active ventilation floor fan, finally generating the optimal PWM signal duty ratio value according to the temperature distribution of the air inlet of the rack, and adjusting the rotating speed of the active ventilation floor fan, so that the temperature distribution of the air inlet of the rack is uniform, and the hot spot problem of the rack is relieved.
2. The intelligent control method for the active ventilation floor of the data center according to claim 1, wherein in the step 2, an optimal Q function is obtained through calculation, that is, an optimal behavior can be selected according to the optimal Q function under any time t system state, so that the accumulated reward is maximized, and the optimal Q function calculation formula is as follows:
Figure FDA0002508954990000031
at any time t, the optimal behavior selection formula is as follows:
Figure FDA0002508954990000032
wherein Q*(φt,at) Represents the optimal Q function, phit+1Represents the state of the system at time t +1, and a represents any of all possible actions taken by the system at time t +1, i.e., the action space
Figure FDA0002508954990000033
A certain behavior in (2).
3. The intelligent control method for active ventilation floor of data center according to claim 1, wherein in step 3, model is solved by using basic intelligent algorithm, sample value variant intelligent algorithm and structure variant intelligent algorithm, and accumulated by continuous decision (phi)t,at,Rt+1,φt+1) Training the neural network by the sample record to enable the neural network to approximate a Q function, and then selectingSelecting optimal behavior such that the cumulative reward for the model is maximized, wheret+1Representing the system state at time t + 1.
4. The intelligent control method for the active ventilation floor of the data center according to claim 3, wherein the basic intelligent algorithm is used for approximating a Q function by using two neural networks with the same structure, one is used for approximating the Q sample function, and a Q sample value is calculated, and the method is called targ network; the other is used for approximating a Q prediction function and calculating a Q prediction value, and is called eval network; and calculating the difference between the Q sample value and the Q predicted value by using the sample record, and training and updating the neural network, wherein the Q sample value calculation formula is as follows:
Figure FDA0002508954990000034
in the sample value variant intelligent algorithm, a Q sample value calculation formula is as follows:
Figure FDA0002508954990000041
wherein Qt+1,targetIs a value of Q sample, Rt+1And phit+1Taken from the sample record, Q (phi)t+1,a；θt,target) For the set of Q samples output by the targ network,
Figure FDA0002508954990000042
for the Q sample set output by targ network, let Qeval(φt+1,a；θt,eval) Q sample value, Q, corresponding to the largest behavioreval(φt+1,a；θt,eval) For the Q prediction set output by the eval network, a represents any one of all possible actions taken by the system at the moment t +1, namely, the action space
Figure FDA0002508954990000043
A certain behavior of, thetat,evalFor the eval network parameter set at time t, θt,targetIs tA time targ network parameter set;
the neural network updating mode is as follows:
Figure FDA0002508954990000044
whereint+1Is the difference between the Q sample value and the corresponding Q predicted value, Q (phi)t,at；θt,eval) In the set of Q predictions output for eval network, atCorresponding predicted value of Q, phitAnd atTaken from the sample record, θt+1,evalFor the eval network parameter set at time t +1,
Figure FDA0002508954990000045
is composed of
Figure FDA0002508954990000046
About thetat,evalα is the neural network learning step size, θtargetIs targ network parameter set when time t is an integer multiple of N including 0evalIs the eval network parameter set when the time t is an integer multiple of N including 0.
5. The intelligent control method for the data center active ventilation floor as claimed in claim 4, wherein the structural variant intelligent algorithm uses two neural networks with the same structure, a DN layer is arranged at the second last layer of each neural network, the DN layer is divided into a V section and an A section, the node number of the V section neurons is 1 and represents the system state at the time t, the number of the A section neurons is the number of elements in the action space and represents all actions which can be taken in the system state, and the calculation formula of the DN layer is as follows:
Figure FDA0002508954990000047
wherein, Q (phi)t,at；θt,θt,V,θt,A) Is the final output of the neural network, phitAnd atTaken from the sample record, θtFor time t, a network parameter set theta before DN layer of the neural network of the structure variant intelligent algorithmt,VFor the DN layer V-segment parameter at time t, θt,AFor the DN segment A parameter at time t, V (phi)t；θt,θt,V) For the output of V section, A (phi)t,at；θt,θt,A) Is a in section AtCorresponding output value, A (phi)t,a'；θt,θt,A) For all outputs of segment A, a' represents the state φtIn the following, all actions that the system may take,
Figure FDA0002508954990000051
the number of elements in the behavior space;
and then, training and updating the neural network by adopting a Q sample value calculation and neural network updating mode which is the same as the sample value variant intelligent algorithm.
6. The intelligent control method for the active ventilation floor of the data center according to claim 1, wherein the operation logic of the intelligent control method is as follows:
1: in different control algorithms, different neural networks are constructed and initialized, and targ network parameters are the same as eval network parameters; setting the sample record cache array; setting a reference temperature
Figure FDA0002508954990000054
2: setting an initial time t to be 0, and recording the time recorded by the sample in the cache array as tau; probability of exploration of initial behavior, rate of exploration decreasing with tMinimum probability of explorationmin；
3: randomly selecting behaviors in Z moments and recording (phi) generated at each momentz∈[0,Z),az∈[0,Z),Rz+1∈[0,Z],φz+1∈[0,Z]) Storing the data into a cache array;
4: obtaining the temperature distribution of the air inlet of the initial frame
Figure FDA0002508954990000052
5: starting a circulation body;
6: obtaining p historical frame air inlet temperature distributions to jointly form a system state phit＝{st-p,…,st-1,st}；
7: if t is equal to 0, then the action a is selectedtMax (dc) and go 9, otherwise go 8;
8: behavior is selected using the following formula:
Figure FDA0002508954990000053
9: execution of atThe PC sends a duty ratio instruction to the microcontroller, changes the rotating speed of the fan and obtains the temperature distribution s of the air inlet of the rack at the next moment of the systemt+1Calculating R according to the winning incentive formula of claim 4t+1；
10: forming a next state phi according to the latest p pieces of temperature distribution historyt+1＝{st+1-p,…,st,st+1And will be (phi)t,at,Rt+1,φt+1) Storing the data into a cache array;
11: randomly extracting Y sample records (phi) from the buffer arrayτ,aτ,Rτ+1,φτ+1)；
12: and calculating the Q sample value by utilizing Y records according to different control algorithms, wherein the formula is as follows:
Figure FDA0002508954990000061
13: updating the eval network using the learning step size α and the loss function:
Figure FDA0002508954990000062
14: exploration probability of-DeltaAndminminimum value of (1);
15: if t mod N is 0, the targ network copies eval network parameters, otherwise 16 is carried out;
16: the time t is increased by 1;
17: the cycle body is ended.
CN202010455152.6A 2020-05-26 2020-05-26 Intelligent control method for data center active ventilation floor Active CN111637614B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010455152.6A CN111637614B (en) 2020-05-26 2020-05-26 Intelligent control method for data center active ventilation floor

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010455152.6A CN111637614B (en) 2020-05-26 2020-05-26 Intelligent control method for data center active ventilation floor

Publications (2)

Publication Number Publication Date
CN111637614A true CN111637614A (en) 2020-09-08
CN111637614B CN111637614B (en) 2021-06-08

Family

ID=72329604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010455152.6A Active CN111637614B (en) 2020-05-26 2020-05-26 Intelligent control method for data center active ventilation floor

Country Status (1)

Country Link
CN (1) CN111637614B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117588394A (en) * 2024-01-18 2024-02-23 华土木(厦门)科技有限公司 AIoT-based intelligent linkage control method and system for vacuum pump

Citations (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030050003A1 (en) * 2001-09-07 2003-03-13 International Business Machines Corporation Air flow management system for an internet data center
CN101634480A (en) * 2009-08-05 2010-01-27 于郡东 System and method for controlling under-floor air distribution air-conditioning fan device in machine room
CN201854532U (en) * 2010-05-17 2011-06-01 罗达俊 Temperature control unit and automatic floor pressurization system with same
CN203024346U (en) * 2013-01-23 2013-06-26 东北石油大学 Air conditioner control system of machine room
CN103292426A (en) * 2012-02-27 2013-09-11 华为技术有限公司 Machine room cooling device and cooling air supply adjusting method
CN103631351A (en) * 2013-12-17 2014-03-12 北京百度网讯科技有限公司 Fan control method and device of server and server
CN203720764U (en) * 2013-12-30 2014-07-16 国家计算机网络与信息安全管理中心 Cooling device used for closed-door cabinet
CN204902119U (en) * 2015-08-25 2015-12-23 贵州电网公司信息通信分公司 Computer room of data center temperature intelligent control device
CN105975029A (en) * 2016-06-13 2016-09-28 天津欧迈通信技术有限公司 Distributed temperature control case radiating system
CN106461260A (en) * 2014-05-15 2017-02-22 三星电子株式会社 Method and apparatus for controlling temperature
CN106528941A (en) * 2016-10-13 2017-03-22 内蒙古工业大学 Data center energy consumption optimization resource control algorithm under server average temperature constraint
CN206283772U (en) * 2016-12-27 2017-06-27 贵州电网有限责任公司信息中心 A kind of active air-flow drainage ventilation unit
CN108446783A (en) * 2018-01-29 2018-08-24 杭州电子科技大学 A kind of prediction of new fan operation power and monitoring method
CN108921223A (en) * 2018-07-05 2018-11-30 广东水利电力职业技术学院（广东省水利电力技工学校） A kind of server cooling system and control method, computer program, computer
CN109654674A (en) * 2018-12-11 2019-04-19 珠海格力电器股份有限公司 Air conditioning system fan control method, air conditioning system and computer readable storage medium
CN110528815A (en) * 2019-06-26 2019-12-03 中电万维信息技术有限责任公司 A kind of based on the controllable anti-static ventilation of the novel Wind Volume of data center machine room Slab
CN110831407A (en) * 2019-11-06 2020-02-21 上海理工大学 Tower type porous air supply temperature control device for data machine room
CN111096094A (en) * 2017-09-06 2020-05-01 维谛公司 Cooling unit energy optimization via intelligent supply air temperature setpoint control
CN111126605A (en) * 2020-02-13 2020-05-08 创新奇智(重庆)科技有限公司 Data center machine room control method and device based on reinforcement learning algorithm
CN111144793A (en) * 2020-01-03 2020-05-12 南京邮电大学 Commercial building HVAC control method based on multi-agent deep reinforcement learning
CN111351180A (en) * 2020-03-06 2020-06-30 上海外高桥万国数据科技发展有限公司 System and method for realizing energy conservation and temperature control of data center by applying artificial intelligence

Patent Citations (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20030050003A1 (en) * 2001-09-07 2003-03-13 International Business Machines Corporation Air flow management system for an internet data center
CN101634480A (en) * 2009-08-05 2010-01-27 于郡东 System and method for controlling under-floor air distribution air-conditioning fan device in machine room
CN201854532U (en) * 2010-05-17 2011-06-01 罗达俊 Temperature control unit and automatic floor pressurization system with same
CN103292426A (en) * 2012-02-27 2013-09-11 华为技术有限公司 Machine room cooling device and cooling air supply adjusting method
CN203024346U (en) * 2013-01-23 2013-06-26 东北石油大学 Air conditioner control system of machine room
CN103631351A (en) * 2013-12-17 2014-03-12 北京百度网讯科技有限公司 Fan control method and device of server and server
CN203720764U (en) * 2013-12-30 2014-07-16 国家计算机网络与信息安全管理中心 Cooling device used for closed-door cabinet
CN106461260A (en) * 2014-05-15 2017-02-22 三星电子株式会社 Method and apparatus for controlling temperature
CN204902119U (en) * 2015-08-25 2015-12-23 贵州电网公司信息通信分公司 Computer room of data center temperature intelligent control device
CN105975029A (en) * 2016-06-13 2016-09-28 天津欧迈通信技术有限公司 Distributed temperature control case radiating system
CN106528941A (en) * 2016-10-13 2017-03-22 内蒙古工业大学 Data center energy consumption optimization resource control algorithm under server average temperature constraint
CN206283772U (en) * 2016-12-27 2017-06-27 贵州电网有限责任公司信息中心 A kind of active air-flow drainage ventilation unit
CN111096094A (en) * 2017-09-06 2020-05-01 维谛公司 Cooling unit energy optimization via intelligent supply air temperature setpoint control
CN108446783A (en) * 2018-01-29 2018-08-24 杭州电子科技大学 A kind of prediction of new fan operation power and monitoring method
CN108921223A (en) * 2018-07-05 2018-11-30 广东水利电力职业技术学院（广东省水利电力技工学校） A kind of server cooling system and control method, computer program, computer
CN109654674A (en) * 2018-12-11 2019-04-19 珠海格力电器股份有限公司 Air conditioning system fan control method, air conditioning system and computer readable storage medium
CN110528815A (en) * 2019-06-26 2019-12-03 中电万维信息技术有限责任公司 A kind of based on the controllable anti-static ventilation of the novel Wind Volume of data center machine room Slab
CN110831407A (en) * 2019-11-06 2020-02-21 上海理工大学 Tower type porous air supply temperature control device for data machine room
CN111144793A (en) * 2020-01-03 2020-05-12 南京邮电大学 Commercial building HVAC control method based on multi-agent deep reinforcement learning
CN111126605A (en) * 2020-02-13 2020-05-08 创新奇智(重庆)科技有限公司 Data center machine room control method and device based on reinforcement learning algorithm
CN111351180A (en) * 2020-03-06 2020-06-30 上海外高桥万国数据科技发展有限公司 System and method for realizing energy conservation and temperature control of data center by applying artificial intelligence

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JIANXIONG WAN等: "Air Flow Measurement and Management for", 《IEEE ACCESS》 *
李永利等: "基于机器学习的数据中心主动地板模型研究", 《计算机仿真》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117588394A (en) * 2024-01-18 2024-02-23 华土木(厦门)科技有限公司 AIoT-based intelligent linkage control method and system for vacuum pump
CN117588394B (en) * 2024-01-18 2024-04-05 华土木(厦门)科技有限公司 AIoT-based intelligent linkage control method and system for vacuum pump

Also Published As

Publication number Publication date
CN111637614B (en) 2021-06-08

Similar Documents

Publication Publication Date Title
CN109189190B (en) Data center heat management method based on temperature prediction
CN111126605B (en) Data center machine room control method and device based on reinforcement learning algorithm
CN100551218C (en) CRAC unit controls based on recycle index
JP6313217B2 (en) Temperature control of HVAC unit
JP2019196898A (en) Building device with cost target optimization
US20150100165A1 (en) Accuracy-optimal control decisions for systems
US8914155B1 (en) Controlling fluid flow in a data center
CN105353611A (en) Refrigeration power balance control method of refrigeration container ship and control system thereof
CN102620378A (en) Method and system for data center energy saving controlling
CN116991215A (en) Causal learning-based data center foundation structure optimization method
CN110618740A (en) Dynamic temperature control device and dynamic temperature control method
CN113885609B (en) Method and device for controlling temperature of refrigerator body of vehicle-mounted refrigerator and vehicle-mounted refrigerator
CN111952692B (en) Automobile battery pack temperature balance control system and control method thereof
CN114279042B (en) Central air conditioner control method based on multi-agent deep reinforcement learning
CN111637614B (en) Intelligent control method for data center active ventilation floor
CN116501104A (en) Temperature control device and method for liquid cooling heat dissipation system
CN113821903A (en) Temperature control method and device, modular data center and storage medium
CN111601490B (en) Reinforced learning control method for data center active ventilation floor
CN117930647A (en) Data center cooling control method and device based on thermal prediction model
CN116954329A (en) Method, device, equipment, medium and program product for regulating state of refrigeration system
CN115717758A (en) Indoor space temperature and humidity regulation and control method and system
CN118093301A (en) Method and device for regulating temperature of server cluster
CN115163540A (en) Fan rotating speed control method, device and equipment
CN115003140A (en) Cooperative control energy-saving method for air conditioner at tail end of water cooling unit of data center machine room
JP2019527328A (en) HVAC system using model predictive control with distributed low-dimensional air supply side 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