Receding Horizon Control of Building Chilled Water Plant – An Integrated Optimization Approach for Chiller Sequencing and Temperature Setpoint Optimization

Abstract

Multi-chiller central cooling plants offer various advantages such as design flexibility and adaptability to changing cooling loads over time. However, due to the sheer complexity of these chiller plants there is a significant potential for energy savings which the existing rule-based control algorithms are unable to explore as they are designed to satisfy minimum energy code requirements. Dynamic scheduling and optimization techniques offer a promising solution to improve the performance of multi-chiller central cooling plants. These algorithms consider various inputs such as weather forecasts, building cooling load, and the performance characteristics of the individual chillers in order to select the most appropriate chiller for a given cooling load. This study develops an integrated optimization algorithm that considers both chiller sequencing (i.e., determining which chillers should be operated at any given time) and temperature setpoint optimization (i.e., determining the optimal temperature setpoints for the chiller plant operation). A deep neural network ensemble model is developed that predicts the future cooling demand for the chiller plant. The power consumption of the chillers and cooling towers are predicted using empirical models which are developed using reference performance curve libraries and calibrated using measured data. These models along with power minimization cost function and temperature/load constraints are combined to formulate the optimization problem as a mixed integer non-linear programming (MINLP) problem. A decomposition algorithm is then used to transform the MINLP optimization problem into a master problem that handles the integer variables and a subproblem that handles the non-linear variables. A graph structure is defined over the finite horizon with the variables and constraints of the problem being represented as nodes and edges respectively and the optimal solution is identified using a graph traverse algorithm. This optimization algorithm is implemented in a receding horizon control framework. This is done to make the actions more robust towards disturbances and to improve the efficiency of the optimization algorithm.

The performance of the proposed framework is tested using Texas A&M University’s Health Science Center’s (HSC) central cooling plant as the baseline. The study predicted an improvement of 15.9% energy savings as compared to the existing rule-based strategy. The energy savings opportunity increases to 17.1% during periods when the plant is operating at ten percent or more of its nominal capacity. In addition to the reduction in power consumption, the proposed algorithm reduced the excess operating capacity of the plant by 31.3%, eliminated the short cycling of the chillers and cooling towers, and improved stability in temperature setpoint resets. However, analyzing the results pertaining to the model performance, a few areas of further analysis are identified such as actual demand satisfaction capabilities which can be made more robust.