Gradient Boosting-Based Residual Correction Model for Occupancy Estimation Using Channel State Information Signals in Under-Actuated Zones

Abstract

This study proposes a robust two-stage framework for estimating occupancy in under-actuated indoor zones using WiFi Channel State Information as a non-intrusive sensing modality. These zones, marked by sparse sensor coverage and irregular airflow, present challenges for conventional detection methods. To address these, a domain-specific feature engineering pipeline is introduced, incorporating temporal, nonlinear, and spatial interaction features derived from filtered channel state information amplitudes. The proposed Gradient Boosting-Based Residual Correction Model combines a Gradient Boosting Regressor to learn primary occupancy patterns with a Histogram-Based Regressor to model structured residual errors. The model was evaluated using two distinct datasets: a controlled dataset with predefined occupancy levels, and an uncontrolled dataset collected in real-world public transit settings with high variability across zones. On the controlled dataset, the proposed model achieved high predictive accuracy with an RMSE of 0.1292, MAE of 0.0952, and R2 of 0.9918. On the uncontrolled dataset, the model retained strong performance with an RMSE of 1.3270, MAE of 1.0933, and R2 of 0.8199, preserving over 82.6% of its accuracy relative to the controlled setting. Ablation studies confirmed the critical role of multipath-aware smoothing features, where their removal led to a +426% increase in MAE. Temporal and spatial features contributed modestly but supported overall robustness. The proposed model offers a scalable, generalizable solution for real-time occupancy monitoring and intelligent HVAC control in both structured and dynamic indoor environments.