國立虎尾科技大學 |

Making Reinforcement Learning Practical for Building Energy Management.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Making Reinforcement Learning Practical for Building Energy Management./
Author:	Dey, Sourav.
Description:	1 online resource (246 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
Contained By:	Dissertations Abstracts International85-06B.
Subject:	Architectural engineering. -
Online resource:	click for full text (PQDT)
ISBN:	9798381164954

Making Reinforcement Learning Practical for Building Energy Management.
Dey, Sourav.

Making Reinforcement Learning Practical for Building Energy Management. - 1 online resource (246 pages)

Source: Dissertations Abstracts International, Volume: 85-06, Section: B.

Thesis (Ph.D.)--University of Colorado at Boulder, 2023.

Includes bibliographical references

Conventional building controls rely on pre-programmed logic and fixed control strategies. Thus, they are typically not flexible and often not adaptive to real-time external and internal conditions. Moreover, with the incorporation of emerging technologies, including solar photovoltaics, electric vehicles, smart devices, Internet of Things (IoT) devices and sensors in buildings, operational control objectives are becoming increasingly complex, calling for advanced controls approaches. Reinforcement learning (RL) is an emerging model-free control method that has the ability to handle complex competing objectives and learning to continuously improve performance by repeatedly interacting with the controlled system in its environment. Although RL is an effective control strategy that can self-learn and adapt to changes in an environment without extensive modeling efforts, it can take an unacceptably long time to learn an effective control strategy. Additionally, the RL controller is unstable in its initial interactions, where it has insufficient system behavior knowledge. This process of trial-and-error and long training time make it unsuitable for an RL controller to be applied directly to a building for direct application.The challenges of rapid learning and early unstable behavior were addressed through the evaluation of various methodologies, including the utilization of surrogate models. Strategies such as imitation learning and offline learning from historical rule-based data were juxtaposed against metamodel training, inverse reinforcement learning, and online learning complemented by guided exploration. It was discovered that an synthesis of online learning and imitation learning was adept at decreasing the training duration and curtailing early exploration. This was achieved without the need for detailed and costly building models and could be executed exclusively through data-driven methods. Nonetheless, the computational cost of this method was found to be substantial. Moreover, the challenge of transfer learning in buildings was tackled to foster scalability of data-driven control approaches. A novel method of rule extraction was explored, which was coupled with the capability of a human stakeholder to interpret, engage with, and adapt the learned policies from one trained agent to another, targeting a specific building of interest.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798381164954Subjects--Topical Terms:

1180400
Architectural engineering.
Subjects--Index Terms:

Conventional building controlsIndex Terms--Genre/Form:

554714
Electronic books.

Making Reinforcement Learning Practical for Building Energy Management.
LDR:03791ntm a22003977 4500 001 1149676
005 20241022112612.5
006 m o d
007 cr bn ---uuuuu
008 250605s2023 xx obm 000 0 eng d
020 $a 9798381164954
035 $a (MiAaPQ)AAI30811904
035 $a AAI30811904
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Dey, Sourav. $3 1476002
245 1 0 $a Making Reinforcement Learning Practical for Building Energy Management.
264 0 $c 2023
300 $a 1 online resource (246 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
500 $a Advisor: Henze, Gregor P.
502 $a Thesis (Ph.D.)--University of Colorado at Boulder, 2023.
504 $a Includes bibliographical references
520 $a Conventional building controls rely on pre-programmed logic and fixed control strategies. Thus, they are typically not flexible and often not adaptive to real-time external and internal conditions. Moreover, with the incorporation of emerging technologies, including solar photovoltaics, electric vehicles, smart devices, Internet of Things (IoT) devices and sensors in buildings, operational control objectives are becoming increasingly complex, calling for advanced controls approaches. Reinforcement learning (RL) is an emerging model-free control method that has the ability to handle complex competing objectives and learning to continuously improve performance by repeatedly interacting with the controlled system in its environment. Although RL is an effective control strategy that can self-learn and adapt to changes in an environment without extensive modeling efforts, it can take an unacceptably long time to learn an effective control strategy. Additionally, the RL controller is unstable in its initial interactions, where it has insufficient system behavior knowledge. This process of trial-and-error and long training time make it unsuitable for an RL controller to be applied directly to a building for direct application.The challenges of rapid learning and early unstable behavior were addressed through the evaluation of various methodologies, including the utilization of surrogate models. Strategies such as imitation learning and offline learning from historical rule-based data were juxtaposed against metamodel training, inverse reinforcement learning, and online learning complemented by guided exploration. It was discovered that an synthesis of online learning and imitation learning was adept at decreasing the training duration and curtailing early exploration. This was achieved without the need for detailed and costly building models and could be executed exclusively through data-driven methods. Nonetheless, the computational cost of this method was found to be substantial. Moreover, the challenge of transfer learning in buildings was tackled to foster scalability of data-driven control approaches. A novel method of rule extraction was explored, which was coupled with the capability of a human stakeholder to interpret, engage with, and adapt the learned policies from one trained agent to another, targeting a specific building of interest.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Architectural engineering. $3 1180400
650 4 $a Computer science. $3 573171
653 $a Conventional building controls
653 $a Machine learning
653 $a Reinforcement learning
653 $a Effective control strategy
653 $a Online learning
655 7 $a Electronic books. $2 local $3 554714
690 $a 0800
690 $a 0462
690 $a 0984
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a University of Colorado at Boulder. $b Civil, Environmental, and Architectural Engineering. $3 1190448
773 0 $t Dissertations Abstracts International $g 85-06B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30811904 $z click for full text (PQDT)