Machine Learning Thermal Control for Data Centers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for thermal control optimization in Internet Data Centers (IDCs) are labor-intensive and costly, often conducted infrequently, which can lead to missed opportunities for ongoing thermal control improvements and incomplete modeling of extraordinary or emergent conditions.
Innovation Solution
A monitoring/control mechanism that uses a combination of Computational Fluid Dynamics (CFD) simulations and machine learning algorithms, such as Convolutional Neural Networks (CNN), to perform real-time, multi-dimensional thermal control optimization. This mechanism collects data from IDC systems, conducts simulations, trains control systems, and updates them based on feedback to continuously optimize thermal control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If CFD analysis is conducted occasionally (e.g., once a year), then the analysis cost is reduced, but thermal control optimization is lost and modeling of extraordinary conditions becomes incomplete
Solution Approach 1:
The patent replaces the traditional mechanical CFD simulation process with a machine learning-based prediction system. The ML model, trained on historical CFD data and thermal control information, can rapidly predict thermal conditions without requiring expensive and time-consuming CFD simulations for each scenario, enabling real-time optimization while reducing computational costs
Solution Approach 2:
The system performs preliminary actions by conducting CFD analysis only when necessary (e.g., when extraordinary conditions occur or periodically) to train and update the ML model, rather than continuously running expensive simulations. This allows the system to maintain optimization capabilities while reducing overall computational resource consumption
2Reliability
If CFD analysis is conducted frequently for real-time optimization, then thermal control optimization is improved, but the analysis cost and complexity increase
Solution Approach 1:
The patent substitutes expensive CFD simulations with a trained machine learning model that can rapidly predict thermal conditions. The ML model, once trained on historical CFD data, provides real-time predictions at minimal computational cost, enabling frequent analysis without the associated expense and time consumption of repeated CFD simulations
3Loss of energy
If CFD analysis is conducted occasionally, then computational resources are saved, but the ability to respond to extraordinary or emergent conditions is lost
Solution Approach 1:
The system performs preliminary training of the ML model using historical CFD data and thermal control information before extraordinary conditions occur. This pre-trained model can then rapidly respond to extraordinary or emergent conditions without requiring time-consuming CFD simulations, maintaining adaptability while reducing computational resource consumption
Solution Approach 2:
The system incorporates feedback mechanisms where actual thermal control outcomes and extraordinary condition responses are fed back into the ML model for continuous learning and improvement. This feedback loop enables the model to adapt to new patterns and conditions over time, maintaining high adaptability without requiring continuous expensive CFD analysis
Data Source
AI summary
Apparatus and methods are provided for improving thermal control, including collecting data of a plurality of systems, each of the plurality of systems including at least one first cooling element and at least one first heat-generating element; conducting a first simulation using a simulation model based on the collected data to generate a first set of simulation results; conducting a first training on a control system using the first set of simulation results to obtain a first trained control system; and using the first trained control system to monitor a field system with a space having at least one second cooling element and at least one second heat-generating element and to control the at least one second cooling element and the at least one second heat-generating element.