A Proposed Mamba Network for Multi-Horizon Prediction of Occupant Air-Conditioning Behavior

Understanding how heating, ventilation, and air-conditioning (HVAC) operating-state predictability and predictive information change with forecast horizon is important for occupant-centric building operation. To systematically examine these changes from near-term to day-ahead horizons, this study proposes a Mamba-based framework for multi-horizon HVAC operating-state prediction. The model integrates occupant-related, environmental, temporal, and historical HVAC information, and combines long-term state-space modeling with recent-operation attention and periodic pattern learning to capture HVAC operating dynamics across different prediction horizons. The model was evaluated on three datasets from the ASHRAE Global Occupant Behavior Database at six prediction horizons from 10 min to 24 h. Prediction performance generally declined with increasing horizon, with larger degradation under sparse and highly imbalanced operating conditions. Feature ablation indicated a shift from strong dependence on the current HVAC state at short horizons toward greater reliance on accumulated operation history and recurring temporal patterns at longer horizons. Compared with ten baseline models spanning conventional machine learning and deep sequence models, the proposed model consistently achieved the highest F1-score across all datasets and prediction horizons, with average relative improvements of 4.7%, 2.8%, and 1.6% in F1-score, PR-AUC, and AUROC, respectively. These findings characterize horizon-dependent prediction performance and feature dependence across the evaluated datasets and inform the selection of forecasting inputs for different operational horizons.