Estimating air exchange rates in thousands of elementary school classrooms using commercial CO2 sensors and machine learning

Abstract

In the post-COVID-19 pandemic era, maintaining clean air in school classrooms has become critical for ensuring student health and safety. Air exchange rate (AER), which measures the number of air replacements in a room per hour, is a standard metric for assessing ventilation, with recommended targets provided by organizations worldwide. Installing comprehensive carbon dioxide (CO₂) monitoring in schools has expanded opportunities for automating AER estimation, but most schools have limited computational resources and analytical capacity. To address this, we developed a cost-effective and scalable method to estimate AER by leveraging end-of-school day carbon dioxide concentrations recorded with thousands of commercial sensors in classrooms. This method assumes well-mixed conditions and replicates the tracer gas technique, leveraging statistical machine learning and knowledge of classroom operations to automate AER calculations at the end of occupied periods. We analyzed data from 3206 sensors across 125 schools in a large urban school district in the Northeastern United States and identified 648,956 CO₂ decay curves over one school year. After applying data screening criteria, we calculated 323,776 AER values, averaging 84 values (SD = 40) per classroom. Calculated AER ranged from < 0.1–64 h⁻¹, averaging 3.0 h⁻¹ (SD = 2.9). The average AER in schools with central mechanical ventilation was 1.8 times higher than in schools without mechanical ventilation. The method is optimized to use parallel and high-performance computing resources, and calculates daily air exchange rates for an entire classroom over an academic school year in a few seconds, an entire school in a few minutes, and the entire school district in a few hours. To our knowledge, this is the largest deployment of commercial CO₂ sensors in schools that publicly share data. The AER calculation method is scalable and efficient, and automates cleaning, selection, and processing of CO₂ data from commercial sensors, with methods and code transferable to other schools collecting similar large-scale data.