Data center micromodule inter-row air conditioner joint optimization scheduling method and system

Abstract

The invention discloses a data center micromodule inter-row air conditioner joint optimization scheduling method and system, and the method comprises the steps: collecting the temperature and wind speed of an air outlet of each inter-row air conditioner at the same moment and the temperature data of a rear measurement point of each IT equipment cabinet in a single micromodule of a data center, and constructing a data set; a temperature prediction model of the incidence relation of the inter-row air conditioner air outlet temperature, the air speed and the case rear measuring point temperature is built; training a temperature prediction model by using the training set, and optimizing intermediate vector dimensions output by an LSTM layer and an Attention layer to obtain a trained temperature prediction model; the inter-row air conditioner air outlet temperature and the air speed set value are adjusted, the temperature prediction model is input, and an inter-row air conditioner joint optimization dispatching value is formed based on the inter-row air conditioner air outlet temperature and the air speed set value; according to the invention, the method of association relation mining and joint optimization scheduling is adopted, and the temperature balance of the IT equipment cabinet measuring points in the data center micro-module under the condition of load change is rapidly realized.

Description

Technical Field

[0001] The present invention relates to a method and system for joint optimization scheduling of row air conditioners, and in particular to a method and system for joint optimization scheduling of micro-module row air conditioners in a data center, belonging to the technical field of precision air conditioner intelligent control. Background Art

[0002] The data center micromodule with closed cold aisles is the solution adopted by the current data center central air-cooled terminal system to deal with the disordered airflow organization of the inter-row air conditioners in the data room. This solution still needs to solve the problem of temperature imbalance at the measuring points of the IT equipment cabinets in the micromodule through long-term manual testing. In order to ensure the normal operation of the equipment in the IT equipment cabinet, the dynamic environmental monitoring system requires that the temperature measuring points of all IT equipment cabinets in the micromodule do not exceed the standard. Due to the imbalance of the temperature of the measuring points of the IT equipment cabinet, the outlet wind speed and temperature of the inter-row air conditioner need to be adjusted according to the highest measuring point temperature of the IT equipment cabinet, resulting in low temperatures at the remaining measuring points. The measuring point temperature is closely related to the energy consumption of the central air-conditioning system. The lower the measuring point temperature, the greater the system energy consumption. Therefore, seeking a method for controlling the joint optimization scheduling of the inter-row air conditioners to achieve a balanced temperature of all IT equipment cabinet measuring points in the micromodule is the problem that the invention intends to solve.

[0003] The existing data center dynamic environment monitoring system only implements single-loop control of a single inter-row air conditioner in a single micromodule. Specifically, the outlet temperature of each inter-row air conditioner is set independently, and the outlet wind speed of each inter-row air conditioner is matched automatically or manually. Since it is difficult to use mechanism analysis to determine the impact of adjusting the outlet temperature of a single inter-row air conditioner on the temperature of different nearby IT equipment cabinet measurement points, the problem of uneven temperature at each measurement point of the IT equipment cabinet in the micromodule can only be solved by manual observation and manual adjustment of the set values ​​of the outlet temperature and outlet wind speed of different inter-row air conditioners. This method has a long adjustment cycle and cannot dynamically respond to changes in equipment load in the IT equipment cabinet. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method and system for joint optimization scheduling of micro-module row air conditioners in a data center that can quickly achieve temperature balance at measurement points in IT equipment cabinets.

[0005] Technical solution: The invention provides a method for joint optimization scheduling of micro-module row air conditioners in a data center, comprising:

[0006] (1) In a single micromodule in the data center, collect the temperature and wind speed of the air outlets of each row of air conditioners and the temperature data of the rear measurement points of each IT equipment cabinet at the same time, build a data set, and select the training set, validation set, and test set from the data set;

[0007] (2) Construct a temperature prediction model that correlates the air outlet temperature and wind speed of the inter-row air conditioner with the temperature of the rear measuring point of the chassis;

[0008] (3) Use the training set to train the temperature prediction model, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model;

[0009] (4) In response to the temperature changes at some measuring points in the IT equipment cabinet, the outlet temperature and wind speed setting values ​​of the inter-row air conditioners are adjusted and input into the temperature prediction model. The temperature of the IT equipment cabinet measuring points does not exceed the preset temperature upper limit and the mean square error of the temperature at each measuring point is minimized. Based on the outlet temperature and wind speed setting values ​​of each inter-row air conditioner, the inter-row air conditioner joint optimization scheduling value is formed.

[0010] Furthermore, the step (1) comprises:

[0011] The data set is constructed by the state vectors at each time, where the state vector S at time k is k , the air outlet temperature TMo of each row air conditioner k 、Air outlet speed of air conditioners between rows VMo k And the temperature Tb of the measuring point at the rear of each equipment cabinet k Three row vectors are concatenated:

[0012] S k =TMo k ||VMo k ||Tb k

[0013] Among them, TMo k The elements in include the outlet temperature of each inter-row air conditioner; VMo k The middle element includes the outlet speed of each inter-row air conditioner, Tb k The elements in include the temperature of the rear point of each IT equipment cabinet,

[0014] Take the temperature Tb of each cabinet rear point at time k+1 k+1 As sample S k TagsLab k+1 ,Right now:

[0015] Lab k+1 =Tb k+1 .

[0016] Furthermore, the step (1) of selecting a training set, a validation set, and a test set from the data set includes:

[0017] In the original data set, find the absolute value of the outlet temperature difference between time k and time k+1 of any inter-row air conditioner |TMo k+1 -TMo k |≥0.5, select this sample and the following x samples S k , Sk+1 、…、S k+x , divided into a training set, a validation set, and a test set according to a preset ratio.

[0018] Furthermore, the step (2) includes:

[0019] Let the input sample be the state vector S at time k k , and the hidden vector output by the LSTM layer is denoted as The hidden vector output by the Attention layer is denoted as The predicted result of the measured temperature at the rear of the IT device cabinet output by the fully connected layer is T k+1 , and make this vector consistent with the variables in the label vector Lab k+1 ;

[0020] The loss function L of the neural network model is:

[0021]

[0022] where n is the dimension of the Lab k+1 vector dimension.

[0023] Furthermore, the step (3) includes:

[0024] Let the dimension of S k be m, set the initial value of the vector dimension p of the hidden vector output by the LSTM layer to be 5m, the Attention layer does not participate in training, and at this time Set the calculation result of the model loss function L to l, the comparison value constant c, the comparison threshold to r, 0 < r < 1, and use the Adam gradient optimization algorithm for model training.

[0025] Furthermore, the use of the Adam gradient optimization algorithm for model training includes:

[0026] (31) When (c - l) / c < r, reduce the initial dimension p of the hidden vector , continue training until r ≤ (c - l) / c < 1 when training terminates;

[0027] (32) When l > c, the Attention layer participates in training, and set the initial value of the vector dimension q of the hidden vector to 2m and conduct training;

[0028] (33) When the result l of the model loss function after training satisfies (c - l) / c < r, return to step (31); when the result of the model loss function is l > c, increase the hidden vector of the LSTM layer and the hidden vector of the Attention layer, and the increase amount is m;

[0029] (34) Training is performed until the model loss function result l satisfies (cl) / c <r。

[0030] Furthermore, the step (4) comprises:

[0031] (41) Input the temperature Tb of all IT equipment cabinet rear measurement points at time k k , when the temperature of any of the rear measurement points deviates from the mean When it is greater than the preset threshold, the air outlet temperature setting value of the inter-row air conditioner near the deviation temperature point is adjusted. When the setting value is lowered, otherwise the setting value is increased;

[0032] (42) The adjusted air outlet temperature setting value TMo k ', wind speed setting value VMo k ′ and Tb k Concatenate to form the input vector S k ,Right now:

[0033] S k =TMo k ′||VMo k ′||Tb k

[0034] (43) The vector S k Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+1 k+1 ;

[0035] (44) TMo k ′、VMo k ′ and T k+1 Splice to form the input vector S at time k+1 k+1 ,Right now:

[0036] S k+1 =TMo k ′||VMo k ′||T k+1

[0037] (45) The vector S k+1 Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+2 k+2 ;

[0038] (46) The temperature vector T of the cabinet rear point at time k+3 is further calculated. k+3 ;

[0039] (47) Adjust the air outlet temperature of the air conditioner between different rows according to the preset temperature gradient, and take the mean square error The minimum air outlet temperature of each row air conditioner is used as the output value of joint scheduling.

[0040] Based on the same inventive concept, the present invention also provides a data center micro-module row air conditioner joint optimization scheduling system, comprising:

[0041] The initialization module is used to collect the temperature and wind speed of the air-conditioning outlets between rows and the temperature data of the rear measurement points of each IT equipment cabinet at the same time in a single micromodule of the data center, build a data set, and filter out the training set, validation set and test set from the data set.

[0042] Model building module, building a temperature prediction model of the correlation between the air outlet temperature and wind speed of the inter-row air conditioner and the temperature of the rear measuring point of the chassis;

[0043] The model training module is used to train the temperature prediction model using the training set, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model;

[0044] The optimization module is used to adjust the temperature and wind speed setting values ​​of the inter-row air conditioner outlets according to the temperature changes of some measuring points in the IT equipment cabinet, input the temperature prediction model, and form the inter-row air conditioner joint optimization scheduling value based on the air outlet temperature and wind speed setting values ​​of each inter-row air conditioner, with the temperature of the IT equipment cabinet measuring point not exceeding the preset temperature upper limit and the minimum mean square error of the temperature of each measuring point as the restriction.

[0045] Based on the same inventive concept, the present invention also provides a computing device, including: one or more processors, one or more memories and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the method for joint optimization scheduling of micro-module row air conditioners in a data center according to any one of the above items are implemented.

[0046] Based on the same inventive concept, the present invention also provides a storage medium, which stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the data center micro-module inter-row air conditioning joint optimization scheduling method according to any one of the above items.

[0047] Beneficial effect: Compared with the prior art, the present invention adopts a three-layer neural network training model of LSTM+attention mechanism to mine the sample data association relationship of two types of influencing factors, time series and spatial position. After the temperature and wind speed setting values ​​of the inter-row air-conditioning outlets at different spatial positions in the micro-module of the data center change, it can predict in real time the temperature of the measuring points at different spatial positions of all IT equipment cabinets after a certain time interval, and obtain the combination of the temperature and wind speed setting values ​​of each inter-row air-conditioning outlet when the mean square deviation of the temperature of the rear measuring points of all IT equipment cabinets after a given time interval is minimized by changing the temperature and wind speed setting values ​​of the inter-row air-conditioning outlets in a gradient manner. The combination is fed back to the existing data center dynamic environment monitoring system as the micro-module joint optimization scheduling value, so as to quickly realize the temperature balance of the measuring points of the IT equipment cabinets in the micro-module of the data center under load changes, and achieve the overall energy saving effect of the inter-row air-conditioning in the micro-module when the highest temperature measuring point meets the standard. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0049] Figure 2 Schematic diagram of a temperature prediction model for IT equipment cabinet measurement points according to an embodiment of the present invention;

[0050] Figure 3 The figure is a schematic diagram of the deployment of devices in a micromodule according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0052] Embodiment 1: As attached Figure 1 As shown, the data center micro-module row air conditioner joint optimization scheduling method of this embodiment includes:

[0053] (1) In a single micromodule in the data center, collect the temperature and wind speed of the air outlets of each row of air conditioners and the temperature data of the rear measurement points of each IT equipment cabinet at the same time, build a data set, and select the training set, validation set, and test set from the data set;

[0054] (2) Construct a temperature prediction model that correlates the inter-row air conditioner outlet temperature, wind speed, and the temperature of the rear measurement point of the chassis;

[0055] (3) Use the training set to train the temperature prediction model, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model;

[0056] (4) In response to the temperature changes at some measuring points in the IT equipment cabinet, the outlet temperature and wind speed setting values ​​of the inter-row air conditioners are adjusted and input into the temperature prediction model. The temperature of the IT equipment cabinet measuring points does not exceed the preset temperature upper limit and the mean square error of the temperature at each measuring point is minimized. Based on the outlet temperature and wind speed setting values ​​of each inter-row air conditioner, the inter-row air conditioner joint optimization scheduling value is formed.

[0057] Specifically: In step (1), the single sample and label of the dataset are constructed

[0058] The data set constructed in this embodiment is a time series data set including all air conditioning outlet temperature setting values ​​and wind speed setting values ​​in a single micromodule and the temperature of each IT equipment cabinet rear measurement point. For the time series, k is used to represent a certain moment.

[0059] A single sample in the data set is the state vector S at time k k , which is composed of the air outlet temperature setting value vector (TMo k ), air conditioning wind speed setting value (VMo k ), the temperature of each cabinet rear measuring point (Tb k ) is composed of three row vectors. That is, a single sample in the data set is:

[0060] S k =TMo k ||VMo k ||Tb k

[0061] S k The vector dimension is counted as m.

[0062] Take the temperature of each cabinet rear point at time k+1 (Tb k+1 ) as sample S k TagsLab k+1 ,Right now:

[0063] Lab k+1 =Tb k+1

[0064] Lab k+1 The vector dimension is counted as n.

[0065] Prediction results of temperature measured at the rear of IT cabinet T k+1 , this vector and the label vector Lab k+1 The variables within are consistent.

[0066] Training data set screening:

[0067] In the original data set, find any inter-row air conditioner whose outlet temperature difference between time k and time k+1 is ≥ 0.5℃ (i.e. |TMo k+1 -TMo kSamples with a value of ≥ 0.5) are selected, and this sample and the subsequent 9-time samples S k , S k+1 , …, S k+9 are included in the training set, validation set, and test set.

[0068] In step (2), a temperature prediction model is modeled.

[0069] The present invention proposes a neural network model composed of an LSTM layer, an Attention layer, and a fully connected layer to construct a temperature prediction model for the measurement points of the IT device cabinet.

[0070] Its input sample is the sample S at time k k (the vector dimension is m), and the hidden vector output by the LSTM layer is denoted as (its vector dimension is denoted as p), the hidden vector output by the Attention layer is denoted as (its vector dimension is q), and the predicted result of the measured temperature at the rear of the IT cabinet output by the fully connected layer is T k+1 (its vector dimension is n).

[0071] Model loss function

[0072] In step (3), the model is trained:

[0073] Set the calculation result of the model loss function L to l, its comparison value constant c, and the comparison threshold to r (0 < r < 1).

[0074] Set the hidden vector output by the LSTM layer The initial value of the vector dimension p is 5m, and the Attention layer does not participate in the training first, that is

[0075] The Adam gradient optimization algorithm is used for model training.

[0076] a. When (c - l) / c < r, reduce the hidden vector The initial dimension p and continue training until the training terminates when r ≤ (c - l) / c < 1.

[0077] b. When l > c, the Attention layer participates in the training, and its hidden vector The initial vector dimension q is set to 2m and continue training.

[0078] c. If the result l of the model loss function after the training in process b satisfies (c - l) / c < r, execute step a; if the result l of the model loss function > c, at the same time increase the hidden vector of the LSTM layer The hidden vector of the Attention layer by an increment of m.

[0079] d. Follow the above steps to train until the process ends at step a.

[0080] Step (4) includes:

[0081] a. Input the temperature Tb of all cabinet rear measurement points at time k k , when the temperature of any measuring point deviates from the mean When it is greater than the threshold, adjust the air outlet temperature setting value of the inter-row air conditioner near the deviation temperature point. When the value is lowered, decrease the set value, otherwise increase the set value.

[0082] b. Set the adjusted outlet temperature to TMo k ', wind speed setting value VMo k ′ and Tb k Concatenate to form the input vector S k ',Right now:

[0083] S k =TMo k ′||VMo k ′||Tb k

[0084] c. Vector S k ′ Input the trained model and obtain the temperature vector T of the rear point of the cabinet at time k+1 k+1 .

[0085] d. TMo k ′、VMo k ′ and T k+1 Splice to form the input vector S at time k+1 k+1 ,Right now:

[0086] S k+1 =TMo k ′||VMo k ′||T k+1

[0087] e. Vector S k+1 Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+2 k+2 ;

[0088] f. Further calculation to obtain the temperature vector T of the rear point of the cabinet at time k+3 k+3 .

[0089] g. Adjust the air outlet temperature of air conditioners between rows by a gradient of 0.5°C and take the mean square error The minimum air outlet temperature of each row of air conditioners is used as the joint scheduling output value, and the joint optimization scheduling value is returned to the existing data center dynamic environment monitoring system through the data interface.

[0090] Based on the same inventive concept, this embodiment also provides a data center micro-module row air conditioner joint optimization scheduling system, including:

[0091] The initialization module is used to collect the temperature and wind speed of the air outlets of the air conditioners in each row and the temperature data of the rear measurement points of each IT equipment cabinet at the same time in a single micromodule of the data center, build a data set, and filter out the training set, validation set, and test set from the data set;

[0092] Model building module, building a temperature prediction model of the correlation between the air outlet temperature and wind speed of the inter-row air conditioner and the temperature of the rear measuring point of the chassis;

[0093] The model training module is used to train the temperature prediction model using the training set, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model;

[0094] The optimization module is used to adjust the temperature and wind speed setting values ​​of the inter-row air conditioner outlets according to the temperature changes of some measuring points in the IT equipment cabinet, input the temperature prediction model, and form the inter-row air conditioner joint optimization scheduling value based on the temperature and wind speed setting values ​​of the inter-row air conditioner outlets, with the temperature of the IT equipment cabinet measuring points not exceeding the preset temperature upper limit and the minimum mean square error of the temperature of each measuring point as the restriction, and return the joint optimization scheduling value to the existing data center dynamic environment monitoring system through the data interface.

[0095] Based on the same inventive concept, this embodiment also provides a computing device, including: one or more processors, one or more memories, and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the method for joint optimization scheduling of micro-module row air conditioners in a data center according to any one of the above items are implemented.

[0096] Based on the same inventive concept, this embodiment also provides a storage medium, which stores a computer program, and the computer program includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the data center micro-module inter-row air conditioner joint optimization scheduling method according to any of the above items.

[0097] Example 2: In this example, in a micromodule with a closed cold channel in a data center, row A deploys 12 IT equipment cabinets and 2 inter-row air conditioners, and row B deploys 11 IT equipment cabinets and 3 inter-row air conditioners. Figure 3The dynamic link monitoring system can collect the temperature of the rear measurement points of the IT equipment cabinet (three each at the top, middle and bottom), the outlet temperature measurement value of the inter-row air conditioner, and the outlet wind speed measurement value.

[0098] TMo k =[AC05.TMo k ,AC10.TMo k ,BC03.TMo k ,

[0099] Bc08.TMo k ,Bc12.TMo k ]

[0100] VMo k =[AC05.VMo k ,AC10.VMo k ,BC03.VMo k ,

[0101] Bc08.VMo k ,Bc12.VMo k ]

[0102]

[0103] A single sample S in the dataset k =TMo k ||VMo k ||Tb k , whose vector dimension is 79.

[0104] Sample S k The dimension of the label and training result vector is 69.

[0105] According to the change of the outlet air temperature of an inter-row air conditioner being greater than or equal to 0.5°C, a data set of the above parameters in the micro-module is formed at the time of the change, 5 minutes, 10 minutes, 15 minutes, ..., and up to 45 minutes after the change.

[0106] In step (2), a temperature prediction model is established, such as Figure 2 As shown; Set the LSTM layer to output the hidden vector The initial value of the vector dimension p is 5*79=395, and the comparison value constant c=0.1 and the comparison threshold r=0.9 of the calculation result of the model loss function L are determined.

[0107] In step (3), the Adam gradient optimization algorithm is used to train the model based on the data set in step (1). The initial learning rate is set to 1e-4, the batch size is set to 64, and the learning rate is reduced to 1 / 4 of the original after every 2 stages. The training is performed for 100 stages at a time.

[0108] The calculation result of the temperature prediction model loss function L is set to l.

[0109] a. When (0.1-l) / 0.1<0.9, reduce the hidden vector The initial dimension is p, and the training continues until 0.9≤(0.1-l) / 0.1<1 when the training is terminated.

[0110] b. When l>0.1, the Attention layer participates in the training, and its hidden vector The initial value of the vector dimension q is set to 2*79=158, and training continues.

[0111] c. If the model loss function result l after training in process b satisfies (0.1-l) / 0.1<0.9, execute step a; if the model loss function result l>0.1, increase the hidden vector of the LSTM layer at the same time Attention layer hidden vector The dimension of , the increase is 79.

[0112] d. Follow the above steps to train until the process ends at step a.

[0113] Step (4) includes:

[0114] a. Input the temperature Tb of all cabinet rear measurement points at time k k , when the temperature of any measuring point deviates from the mean When the temperature is greater than 0.5℃, adjust the air outlet temperature setting value of the inter-row air conditioner near the deviation temperature point. When the value is lowered, decrease the set value, otherwise increase the set value.

[0115] b. Set the adjusted outlet temperature to TMo k ', wind speed setting value VMo k ′ and Tb k Concatenate to form the input vector S k ',Right now:

[0116] S k =TMo k ′||VMo k ′||Tb k

[0117] c. Vector S k ′ Input the trained model and obtain the temperature vector T of the rear point of the cabinet at time k+1 k+1 ′.

[0118] d. TMo k ′、VMo k ′ and T k+1 Splice to form the input vector S at time k+1k+1 ,Right now:

[0119] S k+1 =TMo k ′||VMo k ′||T k+1

[0120] e. Vector S k+1 Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+2 k+2 ;

[0121] f. The same method is used to obtain the temperature vector T of the rear point of the cabinet at time k+3 k+3 .

[0122] g. Adjust the air outlet temperature of air conditioners between rows by a gradient of 0.5°C and take the mean square error The minimum air outlet temperature of each row of air conditioners is used as the joint scheduling output value, and the joint optimization scheduling value is returned to the existing data center dynamic environment monitoring system through the data interface.

[0123] The core of the present invention is to uniformly consider the delay of temperature change, the influence of the position relationship between the inter-row air conditioner and the IT equipment cabinet on the cabinet measurement point temperature, and construct a three-layer neural network training model with LSTM+attention mechanism (hereinafter referred to as "IT equipment cabinet measurement point temperature prediction model", referred to as "temperature prediction model") to improve the accuracy of the prediction of the correlation between the IT equipment cabinet temperature and the inter-row air conditioner outlet wind speed and temperature. Based on this temperature prediction model, when the temperature of some measurement points in the IT equipment cabinet changes, the set values ​​of the outlet temperature and outlet wind speed of each inter-row air conditioner are jointly adjusted to achieve a balanced temperature of each measurement point in the IT equipment cabinet, and to achieve overall energy saving of the inter-row air conditioner in the micromodule when the highest temperature measurement point meets the standard.

Claims

1. A method for joint optimization scheduling of micro-module row air conditioners in a data center, characterized in that: include: (1) In a single micromodule in the data center, collect the temperature and wind speed of the air outlets of each row of air conditioners and the temperature data of the rear measurement points of each IT equipment cabinet at the same time, build a data set, and select the training set, validation set, and test set from the data set; (2) Construct a temperature prediction model that correlates the air outlet temperature and wind speed of the inter-row air conditioner with the temperature of the rear measuring point of the chassis; (3) Use the training set to train the temperature prediction model, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model; (4) In response to the temperature changes at some measuring points in the IT equipment cabinet, the outlet temperature and wind speed setting values ​​of the inter-row air conditioners are adjusted and input into the temperature prediction model. The temperature of the IT equipment cabinet measuring points does not exceed the preset temperature upper limit and the mean square error of the temperature at each measuring point is minimized. Based on the outlet temperature and wind speed setting values ​​of each inter-row air conditioner, the inter-row air conditioner joint optimization scheduling value is formed.

2. The data center micro-module row air conditioner joint optimization scheduling method according to claim 1 is characterized in that: The step (1) comprises: The data set is constructed by the state vectors at each time, where the state vector S at time k is k , the air outlet temperature TMo of each row air conditioner k 、Air outlet speed of air conditioners between rows VMo k And the temperature Tb of the measuring point at the rear of each equipment cabinet k Three row vectors are concatenated: S k =TMo k ∥VMo k ∥Tb k Among them, TMo k The elements in include the outlet temperature of each inter-row air conditioner; VMo k The middle element includes the outlet speed of each inter-row air conditioner, Tb k The elements in include the temperature of the rear point of each IT equipment cabinet, Take the temperature Tb of each cabinet rear point at time k+1 k+1 As sample S k TagsLab k+1 ,Right now: Lab k+1 =Tb k+1 。 3. The data center micro-module row air conditioner joint optimization scheduling method according to claim 2 is characterized in that: Step (1) of selecting a training set, a validation set, and a test set from the data set includes: In the original data set, find the absolute value of the outlet temperature difference between time k and time k+1 of any inter-row air conditioner |TMo k +1 -TMo k |≥0.5, select this sample and the following x samples S k , S k+1 ,…,S k+x , divided into training set, validation set and test set according to preset proportions.

4. The data center micro-module row air conditioner joint optimization scheduling method according to claim 1 is characterized in that: The step (2) comprises: Let the input sample be the state vector S at time k k , the hidden vector output by the LSTM layer is The hidden vector output by the Attention layer is calculated as The fully connected layer outputs the predicted temperature of the rear of the IT equipment cabinet as T k+1 , let this vector and the label vector Lab k+1 The variables within are consistent; The loss function L of the neural network model is: Where n is Lab k+1 Vector dimensions.

5. The data center micro-module row air conditioner joint optimization scheduling method according to claim 1 is characterized in that: The step (3) comprises: Count the dimension of S k as m, and set the initial value of the vector dimension p of the output hidden vector of the LSTM layer to be 5m. The Attention layer does not participate in training. At this time, set the calculation result of the model loss function L to be l, the comparison value constant c, and the comparison threshold to be r, where 0 < r < 1, and use the Adam gradient optimization algorithm to train the model.

6. The data center micro-module row air conditioner joint optimization scheduling method according to claim 1 is characterized in that: The Adam gradient optimization algorithm is used for model training, including: When (c - l) / c < r, reduce the hidden vector For the initial dimension p, continue training until the training terminates when r ≤ (c - l) / c < 1; (32) When l>c, the Attention layer participates in the training and the hidden vector The initial value of the vector dimension q is set to 2m for training; When the result l of the loss function of the trained model satisfies (c - l) / c < r, return to step (31); when the result of the model loss function is l > c, increase the dimensions of the hidden vectors of the LSTM layer and the hidden vectors of the Attention layer by an increment of m; (34) Training is performed until the model loss function result l satisfies (cl) / c <r。 7. The data center micro-module row air conditioner joint optimization scheduling method according to claim 1 is characterized in that: The step (4) comprises: (41) Input the temperature Tb of all IT equipment cabinet rear measurement points at time k k , when the temperature of any of the rear measurement points deviates from the mean When it is greater than the preset threshold, the air outlet temperature setting value of the inter-row air conditioner near the deviation temperature point is adjusted. When the setting value is lowered, otherwise the setting value is increased; (42) The adjusted air outlet temperature setting value TMo k ', wind speed setting value VMo k ′ and Tb k Concatenate to form the input vector S k ,Right now: S k =TMo k ′||VMo k ′||Tb k (43) The vector S k Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+1 k+1 ; (44) TMo k ′、VMo k ′ and T k+1 Splice to form the input vector S at time k+1 k+1 ,Right now: S k+1 =TMo k ′||VMo k ′||T k+1 (45) The vector S k+1 Input the trained model to obtain the temperature vector T of the rear point of the cabinet at time k+2 k+2 ; (46) The temperature vector T of the cabinet rear point at time k+3 is further calculated. k+3 ; (47) Adjust the air outlet temperature of air conditioners in different rows according to the preset temperature gradient, and take the mean square error The minimum air outlet temperature of each row air conditioner is used as the output value of joint scheduling.

8. A data center micro-module row air conditioner joint optimization scheduling system, characterized in that: include: The initialization module is used to collect the temperature and wind speed of the air outlets of the air conditioners in each row and the temperature data of the rear measurement points of each IT equipment cabinet at the same time in a single micromodule of the data center, build a data set, and filter out the training set, validation set, and test set from the data set; Model building module, building a temperature prediction model of the correlation between the air outlet temperature and wind speed of the inter-row air conditioner and the temperature of the rear measuring point of the chassis; The model training module is used to train the temperature prediction model using the training set, optimize the intermediate vector dimensions output by the LSTM layer and the Attention layer, and obtain the trained temperature prediction model; The optimization module is used to adjust the temperature and wind speed setting values ​​of the inter-row air conditioner outlets according to the temperature changes of some measuring points in the IT equipment cabinet, input the temperature prediction model, and form the inter-row air conditioner joint optimization scheduling value based on the air outlet temperature and wind speed setting values ​​of each inter-row air conditioner, with the temperature of the IT equipment cabinet measuring point not exceeding the preset temperature upper limit and the minimum mean square error of the temperature of each measuring point as the restriction.

9. A computing device, characterized in that include: One or more processors, one or more memories, and one or more programs, wherein the programs are stored in the memories and configured to be executed by the processors, and when the programs are loaded into the processors, the steps of the method for joint optimization scheduling of micro-module row air conditioners in a data center according to any one of claims 1 to 7 are implemented.

10. A storage medium, characterized in that: The storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the steps of the method for joint optimization scheduling of micro-module row air conditioners in a data center according to any one of claims 1 to 7.

Images