Authors - Hasin Mahir, Tahfizul Hasan Zihan, Md. Shirazim Munir, Khondkar Ayaz Rabbani, Rifat Ara Rouf, Ferdows Zahid, Mahady Hasan, Md. Tarek Habib Abstract - Indoor air quality (IAQ) is critical for health, comfort, and cognitive performance in classrooms. Yet, many classrooms in lowermiddle-income countries (e.g., Bangladesh) rely on natural ventilation and lack any warning system for impending air-quality guideline breaches. This study is motivated by the need to maintain healthy learning environments in resource-constrained settings, where students and teachers are often exposed to elevated CO2 and particulate levels. To address this, we present the first comparative evaluation of three forecasting models: Prophet, Random Forest (RF), and Long Short-Term Memory (LSTM), on daily PM2.5 and CO2 time series covering six months of continuous operation in Dhaka, Bangladesh. Our goal is to anticipate IAQ deterioration to enable proactive ventilation or filtration interventions. The nonlinear machine-learning models substantially outperform the Prophet baseline. Random Forest performed best for PM. (RMSE 1.87 μg/m³, R² 0.992), showing its ability to capture complex pollutant dynamics. LSTM excelled at forecasting peak CO2 (RMSE 159.7 ppm, R² 0.693), which is critical for timely interventions. These findings demonstrate the feasibility of accurate, low-cost IAQ forecasting in resource-constrained classrooms and underscore the potential of data-driven forecasting to maintain healthier learning environments.
