Lukas Leindals, Peter Grønning, Dominik Franjo Dominković, Rune Grønborg Junker
{"title":"Context-aware reinforcement learning for cooling operation of data centers with an Aquifer Thermal Energy Storage","authors":"Lukas Leindals, Peter Grønning, Dominik Franjo Dominković, Rune Grønborg Junker","doi":"10.1016/j.egyai.2024.100395","DOIUrl":null,"url":null,"abstract":"<div><p>Data centers are often equipped with multiple cooling units. Here, an aquifer thermal energy storage (ATES) system has shown to be efficient. However, the usage of hot and cold-water wells in the ATES must be balanced for legal and environmental reasons. Reinforcement Learning has been proven to be a useful tool for optimizing the cooling operation at data centers. Nonetheless, since cooling demand changes continuously, balancing the ATES usage on a yearly basis imposes an additional challenge in the form of a delayed reward. To overcome this, we formulate a return decomposition, Cool-RUDDER, which relies on simple domain knowledge and needs no training. We trained a proximal policy optimization agent to keep server temperatures steady while minimizing operational costs. Comparing the Cool-RUDDER reward signal to other ATES-associated rewards, all models kept the server temperatures steady at around 30 °C. An optimal ATES balance was defined to be 0% and a yearly imbalance of −4.9% with a confidence interval of [−6.2, −3.8]% was achieved for the Cool 2.0 reward. This outperformed a baseline ATES-associated reward of 0 at −16.3% with a confidence interval of [−17.1, −15.4]% and all other ATES-associated rewards. However, the improved ATES balance comes with a higher energy consumption cost of 12.5% when comparing the relative cost of the Cool 2.0 reward to the zero reward, resulting in a trade-off. Moreover, the method comes with limited requirements and is applicable to any long-term problem satisfying a linear state-transition system.</p></div>","PeriodicalId":34138,"journal":{"name":"Energy and AI","volume":"17 ","pages":"Article 100395"},"PeriodicalIF":9.6000,"publicationDate":"2024-07-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.sciencedirect.com/science/article/pii/S2666546824000612/pdfft?md5=b17bfa78652179749ed19203f3f51d82&pid=1-s2.0-S2666546824000612-main.pdf","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Energy and AI","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2666546824000612","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Data centers are often equipped with multiple cooling units. Here, an aquifer thermal energy storage (ATES) system has shown to be efficient. However, the usage of hot and cold-water wells in the ATES must be balanced for legal and environmental reasons. Reinforcement Learning has been proven to be a useful tool for optimizing the cooling operation at data centers. Nonetheless, since cooling demand changes continuously, balancing the ATES usage on a yearly basis imposes an additional challenge in the form of a delayed reward. To overcome this, we formulate a return decomposition, Cool-RUDDER, which relies on simple domain knowledge and needs no training. We trained a proximal policy optimization agent to keep server temperatures steady while minimizing operational costs. Comparing the Cool-RUDDER reward signal to other ATES-associated rewards, all models kept the server temperatures steady at around 30 °C. An optimal ATES balance was defined to be 0% and a yearly imbalance of −4.9% with a confidence interval of [−6.2, −3.8]% was achieved for the Cool 2.0 reward. This outperformed a baseline ATES-associated reward of 0 at −16.3% with a confidence interval of [−17.1, −15.4]% and all other ATES-associated rewards. However, the improved ATES balance comes with a higher energy consumption cost of 12.5% when comparing the relative cost of the Cool 2.0 reward to the zero reward, resulting in a trade-off. Moreover, the method comes with limited requirements and is applicable to any long-term problem satisfying a linear state-transition system.