Ensemble modelling of hygrothermal system for multi- objectives model predictive controllers

Abstract

An air conditioner maintains occupants’ thermal comfort; however, it is also power-hungry in which leads to high electricity consumption. Maintaining thermal comfort while minimising electrical power consumption is difficult, especially when the weather-related inputs are not predictable. This leads to overcooling due to undershoot and undercooling due to overshoot. Undercooling causes discomfort, while overcooling causes discomfort and high-power usage. To minimise this problem, the implementation of model-based predictive controllers can be developed to produce necessary pre-emptive control decisions based on the output of the embedded simulation model in the controller. However, the simulation model must be accurate for the best result. This project develops accuracy-improved mathematical models that represent the dynamic hygrothermal behaviour of a laboratory in aiding future potential energy-efficient predictive controllers. This is to maintain the thermal comfort level in the laboratory while minimising power consumption. Two thermal comfort variables were modelled to maintain two different desired setpoints simultaneously in the future. First, the empirical modelling was developed to capture the dynamics of the temperature and humidity behaviours of the laboratory using three existing standard methods, which were: (1) autoregressive–moving-average (ARMA) model; (2) transfer function (TF) model; and (3) nonlinear autoregressive exogenous model (NARX) model. Second, the ensemble methods were implemented to increase the simulation accuracy of the developed modelling by summing up the output values from all three developed models – prior to summation, the output of each of the models was multiplied by the weight value assigned for each of the models. The values of these weights were determined using the following three ensemble methods: (1) weighted average; (2) least square technique (LST) / least square method (LSM); and (3) genetic algorithm (GA). All models’ simulation outputs were compared with the actual data for accuracy benchmarking. Results showed that the most accurate ensemble models have better accuracies than the most accurate individual models developed in this research while being simulated with the testing data set in each simulation case. The improvements in the air temperature simulation models are by 3.40%, 7.38%, and 8.69% each for one-, five-, and ten-minute(s) simulation ahead, while the improvements in the relative humidity simulation models are by 0.96%, 1.35%, and 2.38% each for one-, five-, and ten-minute(s) simulation ahead. The accuracy-improved models can then be utilised in model-based predictive controllers for maintaining occupants’ thermal comfort in a building while minimising the air conditioners’ power consumption for energy saving and environmental conservation.