Optimal control of chiller plants using Bayesian network

Linear regression models trained by the dataset, which contain the optimal values for the control parameters under different operating conditions, have been heavily studied in the literature of chiller plants operation optimization due to their performances with higher speeds. However, the linear regression models face difficulties when a nonlinear input/output relationship is considered. Addressing this challenge, we proposed a Bayesian Network (BN) model (a datadriven and probabilistic graphical model), for the operation optimization of chiller plants. Here, we first introduced the construction of the BN model, and demonstrated its validity on the model predictive control of a condenser water set point for a watercooled chiller plant. Then, we evaluated the performance of the proposed BN model under imperfect prediction of weather conditions and building cooling loads with inputs embedding manually generated errors. The baseline performance was provided through a model-based optimization (MBO) using an exhaustive search. The results show that the proposed BN model could provide energy savings compatible with the one given by the MBO using an exhaustive search for both inputs with and without errors.