BROWSE
Detail
Journal Article
Stochastic model predictive control for chiller systems using machine-learning models with quantified prediction uncertainty
Cited 0 time in 
Share 
Share
- Authors
- Sunghyun Kim, Hansol Shin, Jeeye Mun, Joa Hyoung Lee, Yoonmee Doh, Cheol-Soo Park
- Issue Date
- 2026-10
- Citation
- Energy and Buildings, v.369, Part A, pp.1-13
- ISSN
- 0378-7788
- Publisher
- Elsevier
- Language
- English
- Type
- Journal Article
- Abstract
- Machine learning (ML)-based model predictive control (MPC) has emerged as a promising approach for building-energy optimization. However, ML predictions contain inherent uncertainty. Conventional deterministic MPC (DMPC) treats point predictions as the ground truth, leading to optimistic control decisions that tend to have the potential for constraint violations and unreliable estimates of energy savings. This study proposes a stochastic MPC (SMPC) framework that explicitly incorporates ML prediction uncertainty into both the objective function and constraints, thus enabling uncertainty-aware optimal control. The framework employs a long-short-term-memory–based cooling-load prediction model and a transfer-learning–based hybrid chiller COP model using real-world operational data, with the prediction uncertainty quantified using deep ensembles. It is found that DMPC is likely to select high-uncertainty control decisions, producing unreliable distributions of energy saving. Alternatively, the proposed SMPC consistently shifts the distribution toward higher and more reliable outcomes. The mean energy saving potential reaches 32.2% (this value is 10.8% under DMPC), while the load-constraint violation probability decreases from 20.1% to nearly zero. Additionally, a tunable safety factor provides operators with a transparent mechanism for balancing energy efficiency and operational reliability. The findings demonstrate that the ML prediction uncertainty is not merely a source of error to be minimized; it acts as a structured signal that can be leveraged to improve the reliability and efficiency of data-driven building-energy control.
- Keyword
- Model predictive control, Machine learning, Uncertainty, Stochastic, Chiller system
- KSP Keywords
- Balancing Energy, Chiller system, Constraint violation, Data-Driven, Energy Efficiency(EE), Energy saving potential, Load prediction model, Long Short-Term memory(LSTM), Machine learning (ml), Objective function, Operational Data