Optimizing data center controls using neural networks

• US 10,643,121 B2
• Filed: 01/19/2017
• Issued: 05/05/2020
• Est. Priority Date: 01/19/2017
• Status: Active Grant

• Current Assignee: DeepMind Technologies Limited (Alphabet Inc.)
• Original Assignee: DeepMind Technologies Limited (Alphabet Inc.)
• Inventors: Evans, Richard Andrew, Gao, Jim, Ryan, Michael C., Dulac-Arnold, Gabriel, Scholz, Jonathan Karl, Hester, Todd Andrew
• Primary Examiner(s): Zhen, Li B.
• Assistant Examiner(s): Haedi, Selene A.
• Application Number:
  US15/410,547
• Publication Number:
  US 20180204116A1
• Time in Patent Office: 1,202 Days
• CPC Class Codes: G06N 3/045 Combinations of networks

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for improving operational efficiency within a data center by modeling data center performance and predicting power usage efficiency. An example method receives a state input characterizing a current state of a data center. For each data center setting slate, the state input and the data center setting slate are processed through an ensemble of machine learning models. Each machine learning model is configured to receive and process the state input and the data center setting slate to generate an efficiency score that characterizes a predicted resource efficiency of the data center if the data center settings defined by the data center setting slate are adopted t. The method selects, based on the efficiency scores for the data center setting slates, new values for the data center settings.

Claims

1. A method comprising:

receiving a state input characterizing a current state of a data center;

receiving data identifying a first set of data center setting slates that each define a respective combination of possible data center settings that affect a resource efficiency of the data center;

generating a second set of data center setting slates from the first set of data center setting slates, comprising;

for each data center setting slate in the first set of data center setting slates and each of one or more operating constraints for the data center;

processing the state input and the data center setting slate through one or more respective first machine learning models that are each specific to the operating constraint, wherein each first machine learning model is configured to (i) process the state input and the data center setting slate and (ii) generate a corresponding constraint score that characterizes a predicted value of an operating property of the data center if data center settings defined by the data center setting slate are adopted in response to receiving the state input, and

generating a respective final constraint score for the data center setting slate from one or more respective constraint scores generated by the one or more respective first machine learning models;

removing, for each operating constraint and from the first set of data center setting slates, any data center setting slate having a respective final constraint score corresponding to the operating constraint that does not satisfy a respective threshold;

for each data center setting slate in the second set of data center setting slates;

processing the state input and the data center setting slate through each second machine learning model in an ensemble of second machine learning models that are different than the one or more first machine learning models, wherein each second machine learning model in the ensemble is configured to;

receive the state input and the data center setting slate, and

process the state input and the data center setting slate to generate a respective efficiency score that characterizes a predicted resource efficiency of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input; and

selecting, based on the efficiency scores for the data center setting slates in the second set of data center setting slates, new values for settings of the data center.

2. The method of claim 1, wherein the respective efficiency score is a predicted long-term power usage effectiveness (PUE) of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input.

3. The method of claim 2, wherein the ensemble of second machine learning models have been configured to generate the predicted long-term PUE of the data center through training the ensemble of second machine learning models on training data that includes a plurality of training state inputs and, for each training state input, a target PUE that was the PUE of the data center a predetermined time after the data center was in a state characterized by the training state input.

4. The method of claim 3, further comprising:

receiving raw training state inputs; and

preprocessing the raw training state inputs to generate the training state inputs.

5. The method of claim 3, wherein the predetermined time is greater than thirty minutes.

6. The method of claim 5, wherein the predetermined time is one hour.

7. The method of claim 1, wherein selecting new values for the data center settings comprises:

determining, for each data center setting slate, a respective aggregate resource efficiency score from the respective efficiency scores generated for the data center setting slate by the ensemble of second machine learning models;

determining, for each data center setting slate, a respective measure of variation of the respective efficiency scores generated for the data center setting slate by the ensemble of second machine learning models;

ranking the data center setting slates based on the aggregate resource efficiency scores and the measures of variation; and

selecting a combination of possible data center settings defined by a highest-ranked data center setting slate as the new values for the data center settings.

8. The method of claim 7, wherein an aggregate resource efficiency score for a data center setting slate is a measure of central tendency of the efficiency scores generated by the ensemble of second machine learning models for the data center setting slate.

9. The method of claim 7, wherein ranking the data center setting slates comprises ranking the data center slates in an exploitative manner by penalizing data center setting slates that have higher measures of variation more in the ranking than data center setting slates that have lower measures of variation.

10. The method of claim 7, wherein ranking the data center setting slates comprises ranking the data center slates in an explorative manner by promoting data center setting slates that have higher measures of variation more in the ranking than data center setting slates that have lower measures of variation.

11. The method of claim 1, wherein one of the one or more operating constraints is a constraint on the temperature of the data center over a next hour and the operating property of the data center that corresponds to the operating constraint is the temperature of the data center over the next hour.

12. The method of claim 1, wherein one of the one or more operating constraints is a constraint on the pressure of the data center over a next hour and the operating property of the data center that corresponds to the operating constraint is the pressure of the data center over the next hour.

13. The method of claim 1, wherein each second machine learning model in the ensemble has been trained on a different sample of training data from each other second machine learning model in the ensemble of second machine learning models or has a different model architecture from each other second machine learning model in the ensemble of second machine learning models.

14. The method of claim 1, further comprising:

receiving data identifying a true value of the efficiency score for the data center at a time after the data center was in the current state; and

using the state input, the new values for the data center settings, and the true value of the efficiency score in re-training the ensemble of second machine learning models.

15. A system comprising:

one or more computers; and

one or more storage devices storing instructions that are operable, when executed by one or more computers, to cause the one or more computers to perform operations comprising;

receiving a state input characterizing a current state of a data center;

receiving data identifying a first set of data center setting slates that each define a respective combination of possible data center settings that affect a resource efficiency of the data center;

generating a second set of data center setting slates from the first set of data center setting slates, comprising;

for each data center setting slate in the first set of data center setting slates and each of one or more operating constraints for the data center;

processing the state input and the data center setting slate through one or more respective first machine learning models that are each specific to the operating constraint, wherein each first machine learning model is configured to (i) process the state input and the data center setting slate and (ii) generate a corresponding constraint score that characterizes a predicted value of an operating property of the data center if data center settings defined by the data center setting slate are adopted in response to receiving the state input, and

generating a respective final constraint score for the data center setting slate from one or more respective constraint scores generated by the one or more respective first machine learning models;

removing, for each operating constraint and from the first set of data center setting slates, any data center setting slate having a respective final constraint score corresponding to the operating constraint that does not satisfy a respective threshold;

for each data center setting slate in the second set of data center setting slates;

processing the state input and the data center setting slate through each second machine learning model in an ensemble of second machine learning models that are different than the one or more first machine learning models, wherein each machine learning model in the ensemble is configured to;

receive the state input and the data center setting slate, and

processing the state input and the data center setting slate to generate a respective efficiency score that characterizes a predicted resource efficiency of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input; and

selecting, based on the efficiency scores for the data center setting slates in the second set of data center setting slates, new values for settings of the data center.

16. The system of claim 15, wherein the respective efficiency score is a predicted long-term power usage effectiveness (PUE) of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input.

20. The system of claim 15, wherein the ensemble of second machine learning models have been configured to generate the predicted long-term PUE of the data center through training the ensemble of second machine learning models on training data that includes a plurality of training state inputs and, for each training state input, a target PUE that was the PUE of the data center a predetermined time after the data center was in a state characterized by the training state input.

17. One or more non-transitory computer-readable storage mediums storing instructions that are executable by a processing device and upon such execution cause the processing device to perform operations comprising:

receiving a state input characterizing a current state of a data center;

receiving data identifying a first set of data center setting slates that each define a respective combination of possible data center settings that affect a resource efficiency of the data center;

generating a second set of data center setting slates from the first set of data center setting slates, comprising;

for each data center setting slate in the first set of data center setting slates and each of one or more operating constraints for the data center;

processing the state input and the data center setting slate through one or more respective first machine learning models that are each specific to the operating constraint, wherein each first machine learning model is configured to (i) process the state input and the data center setting slate and (ii) generate a corresponding constraint score that characterizes a predicted value of an operating property of the data center if data center settings defined by the data center setting slate are adopted in response to receiving the state input, and

generating a respective final constraint score for the data center setting slate from one or more respective constraint scores generated by the one or more respective first machine learning models;

removing, for each operating constraint and from the first set of data center setting slates, any data center setting slate having a respective final constraint score corresponding to the operating constraint that does not satisfy a respective threshold;

for each data center setting slate in the second set of data center setting slates;

processing the state input and the data center setting slate through each second machine learning model in an ensemble of second machine learning models that are different than the one or more first machine learning models, wherein each machine learning model in the ensemble is configured to;

receive the state input and the data center setting slate, and

process the state input and the data center setting slate to generate a respective efficiency score that characterizes a predicted resource efficiency of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input; and

selecting, based on the efficiency scores for the data center setting slates in the second set of data center setting slates, new values for settings of the data center.

18. The non-transitory computer-readable storage medium of claim 17, wherein the respective efficiency score is a predicted long-term power usage effectiveness (PUE) of the data center if the data center settings defined by the data center setting slate are adopted in response to receiving the state input.

19. The non-transitory computer-readable storage medium of claim 18, wherein the ensemble of second machine learning models have been configured to generate the predicted long-term PUE of the data center through training the ensemble of second machine learning models on training data that includes a plurality of training state inputs and, for each training state input, a target PUE that was the PUE of the data center a predetermined time after the data center was in a state characterized by the training state input.