Advancing Indoor Air Quality Management Through Predictive Feedforward Control - A Comparative Analysis

In the contemporary built environment, ensuring optimal Indoor Air Quality (IAQ) while maintaining energy efficiency represents a multifaceted challenge for building management systems (BMS). Traditional ventilation strategies are predominantly reactive, adjusting air exchange rates only after pollutant concentrations exceed established safety thresholds. This reactive latency, or “phase lag,” often results in temporary exposure to high pollutant levels and suboptimal energy utilization. This study proposes a proactive “feed-forward” control framework based on pollutant forecasting. We conduct a rigorous comparative analysis of three distinct modeling paradigms: Long Short-Term Memory (LSTM) neural networks, Random Forest (RF) ensembles, and Autoregressive Integrated Moving Average (ARIMA) models. The models were trained and validated using a comprehensive 1-year office environmental dataset from Kaggle, encompassing CO₂, PM_2.5 temperature, and humidity. Our findings demonstrate that for CO₂ forecasting, both LSTM and RF architectures achieve high predictive accuracy (R² > 0.97) and successfully eliminate the reactive phase lag. For PM_2.5, the stochastic nature of indoor particulate events presents a significant challenge, with the LSTM exhibiting a “conservative smoothing” effect and occasional mathematical artifacts such as negative values.