Comparison of the Robustness of RNN, MPC and ANN Controller for Residential Heating System

Abstract

In this paper, a novel random neural network (RNN) controller is proposed to maintain a comfortable indoor environment in a single storey residential building having four rooms fitted with radiators for heating. This controller considers the effect of outside temperature and solar radiations on the building and is capable of maintaining a comfortable indoor environment on the basis of a PMV-based set point. The RNN controller is trained with a 30 day dataset from the living room of the building and the performance of the controller is evaluated by testing the controller in all four rooms of the building for 100 days. It is found that the RNN controller is not only capable of maintaining comfortable indoor environment as suggested by PMV-based set point but can also adjust the room temperature to a lower set point (not included in the training set) required by the user for unoccupied rooms. The RNN controller is further compared with similar artificial neural network (ANN) controller and model predictive control (MPC) controller. The results show that for maintaining comfortable indoor environment, the performance of the RNN controller is approximately equivalent to the MPC controller for the set points not covered in the training set, while ANN controller failed to maintain accurate comfortable environment for the operating points not covered in the training phase.