CO2-Based Occupancy Estimation with Efficient Deep Learning: A Step Towards Edge-Deployable Soft Sensors

Accurate occupancy estimation is a key enabler for intelligent building automation, affecting energy efficiency, HVAC control, and space utilization. While conventional sensors such as cameras or motion detectors are either intrusive or expensive, CO2 concentration offers a non-intrusive and widely available alternative correlated with human presence. However, most CO2-based approaches rely on limited real-world datasets and do not systematically evaluate different model architectures or input feature designs. In this study, we leverage synthetic data generated via EnergyPlus simulations to benchmark lightweight deep learning models for CO2-based occupancy estimation. We explore the effectiveness of incorporating temporal CO2 derivatives into model inputs. Our evaluation includes six model architectures (LSTM, GRU, BiLSTM, TCN, Small MLP, and Separable CNN) assessed under three input configurations across diverse room sizes and occupancy patterns. Results show that adding CO2 derivatives significantly improves performance across all models. GRU achieved the best balance between accuracy (MAE = 0.1048, R² = 0.967) and inference efficiency, while Separable CNN offered a strong lightweight alternative with fast runtime and competitive accuracy. These findings demonstrate the viability of using DL-based ’soft sensors’ for real-time, non-intrusive occupancy estimation in edge environments.