Efficiency and Reliability Joint Optimization of Chiller Plants Based on a Hybrid Model

• Published in: IEEE robotics and automation letters Vol. 4; no. 4; pp. 3224 - 3231
• Main Authors: Zhang, Danxu, Mittal, Khushboo, Wilson, James, Luh, Peter B., Fan, Junqiang, Gupta, Shalabh
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01-10-2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Chill rolls; Component reliability; Computer network reliability; Cooling; Decomposition; deep neural network and mixed-integer nonlinear; dynamic; Dynamic programming; Efficiency; Finite difference method; Joint optimization; Mathematical model; Mathematical models; Optimization; Plant reliability; Power demand; Reliability; Reliability analysis; Taylor series; Time
• ISSN: 2377-3766; 2377-3766
• DOI: 10.1109/LRA.2019.2924126

Methods for chiller plant energy savings may increase component degradation, thereby decreasing reliability. Joint optimization is thus important. The problem, however, is challenging. First, existing reliability models which are functions of time in terms of years are not suitable for operation optimization. Second, efficiency optimization is static and often runs every 10-15 min given current demand while reliability change in a short period is not obvious. In this letter, a dynamic chiller reliability model and a static hybrid plant model consisting of empirical and DNN models are developed, and a weighted sum of 1-h plant power and chiller reliability is minimized with a time interval 10 min. The formulation consisting of six independent efficiency and one dynamic reliability optimization problems, and chiller power and reliability are coupled with some common variables. To address the two-time scale issue, the long time scale reliability is approximated by reliability change as a result of operations using Taylor series. To efficiently solve the problem with dynamics, mixed-integers, nonlinearity, and no explicit equations inDNN, a recently developed decomposition and coordination-based method is combined with dynamic programming with rollout and the plant is decomposed into a simplified dynamic chiller subproblem and three simple static subproblems. Gradients needed are obtained by using finite difference without requiring explicit equations. Numerical testing demonstrates the advantages of joint optimization in terms of energy savings and reliability improvement as compared with a baseline.