Evaluating the Benefits of Indoor Air Quality Forecasting for Controlling Particle Filter Systems

Abstract

Predicting future concentrations of fine particulate matter (PM_2.5) and other indoor air pollutants using machine learning is an increasingly frequent topic of research. Although prediction has several proposed applications, the potential benefit has remained largely unevaluated. This study examines whether prediction can improve how particle filter systems such as portable air cleaners are automated by comparing a model predictive control (MPC) strategy with a traditional threshold-based control (TBC) strategy. The MPC controller is designed to optimize the balance between mean reduction in PM_2.5 concentration (i.e., keep concentrations as low as possible following health-based guidance) and reductions in system runtime. These two parameters are compared for both control strategies using 104 simulations of week-long continuous PM_2.5 measurements in occupied apartments. Our findings suggest that there is no meaningful difference in performance between the two control strategies. Additionally, we find only a marginal improvement in performance for MPC controllers that operate using longer prediction horizons. Given the numerous challenges associated with accurately predicting future PM_2.5 concentrations, as well as the additional challenges associated with implementing MPC, controlling particle filter systems may not be an appropriate application for prediction given the lack of benefit over TBC. We would recommend that particle filter systems continue to be controlled using nonpredictive strategies such as TBC, and that prediction tools be used for other applications or adapted to different areas of indoor air quality research and practice.