Optimal control of ventilation in enclosed spaces using Adversarial Neural Networks

Abstract

Systems that manage and control air quality are the main energy consumers within a building and their design often relies on assumptions that lead to excessive energy usage. This paper proposes a new method based on Computational Fluid Dynamics (CFD) and Artificial Intelligence (AI) that can assimilate observations and control ventilation systems to achieve a certain goal, such as maintaining a particular temperature range in an indoor environment. An AI-based reduced-order model (ROM) is used here because current CFD methods generally have too large a computational expense for real-time control. We use proper orthogonal decomposition (POD) to represent the spatial distributions of the CFD simulations and learn the evolution of the POD coefficients in time with an Adversarial Neural Network. With this AI-based ROM, we can perform 4D Variational Data Assimilation (4D-Var DA) and control with a rapid workflow that incorporates the spatial variation of all the key variables: air flow velocity, CO2, temperature, relative humidity and viral load. The proposed method is applied to three scenarios: (i) a ventilation scenario in which the aim is to keep the occupants healthy (indicated by ventilation and CO2 levels) as well as thermally comfortable while minimising energy consumption; (ii) a pandemic scenario in which the priority is to keep people infection-free; (iii) a compromise between (i) and (ii). These three scenarios are developed within a classroom containing 26 children and one teacher, and are combined with an extensive set of measurement data, collected in the classroom of a primary school located in London. Our method successfully met the control objectives in all three scenarios.