Set-Point Optimizer for the Frontier Central Energy Plant (CEP) (2021 - 2023)
Jun 1, 2021
·
1 min read
Analytics and AI Methods at Scale Group (AAIMS), Oak Ridge National Laboratory, USA
Research Staff, July 2021 ~ March 2023
Research Staff, July 2021 ~ March 2023
Overview
This project developed a machine-learning-based optimization system to assist field engineers in tuning and optimizing the high-temperature water (HTW) cooling system for the Frontier supercomputer’s central energy plant.
- Designed a data-driven optimization system that recommends control parameters to improve cooling efficiency and stability.
- Implemented a long short-term memory (LSTM) model to predict system dynamics and guide optimization.
- Used sequential least squares programming (SLSQP) to search for optimal control parameters.
- Developed an interactive web-based tool allowing field engineers to conduct semi-automated “what-if” experiments.
- Provided initial insights into the effectiveness and limitations of data-driven optimization in mission-critical HPC cooling operations.
Technology
- Facility: Bacnet/IP, custom devices, Metasys historian
- Modeling: Multi-variate LSTM (TensorFlow 2)
- Optimization: SLSQP (SciPy - Optimize)
- Application: Javascript - ReactJS, FastAPI, MLflow
