Neural network and polynomial approximated thermal comfort models for HVAC systems

Nowadays, the majority of people carry on their daily activities inside a building. This has motivated research directed to assure several comfort conditions. Thermal comfort is usually maintained by means of HVAC (Heating, Ventilation and Air Conditioning) systems. The most widely used thermal comfort index is the PMV (Predictive Mean Vote), which is computed considering measurements of several physical variables. The classical calculation of this index is expensive in computational terms, and the involved measurement requires a relatively extensive sensor network. This work proposes the use of two approximated models for the PMV index, one is based on an artificial neural network and the other makes use of polynomial expansions, aimed at using these approximated indices within model predictive control frameworks. In this context, the advantages of using approximated models are two-fold: the computational cost of the calculation of the index is reduced, allowing its use in real-time control of HVAC systems; and the network sensor size is decreased. These advantages entail economic benefits and promote the deployment of comfort controllers in larger structures. This paper illustrates the development of the above cited approximated models and includes experimental tests that rate the accuracy and benefits of the proposed models.