Multi-horizon forecasting of thermal and ventilation risks in naturally ventilated hospitals using multi-output transformers

Conference item

Hospital wards in tropical regions face sustained challenges in maintaining acceptable thermal comfort and indoor air quality, particularly where mechanical cooling and ventilation are not available. This study presents a multihorizon forecasting framework for indoor environmental risks in naturally ventilated healthcare settings. The framework draws on more than two years of 15-minute interval sensor measurements collected from 14 wards across two regional hospitals in Tamale, Ghana, paired with weather data, to forecast indoor air temperature, relative humidity (RH), and carbon dioxide (CO₂) concentration over a 24-hour horizon. We benchmark a Temporal Fusion Transformer (TFT) against a multi-output XGBoost baseline, while current environmental risk is summarised against international standards at ward-level. The risk summary indicates that overheating and elevated humidity are widespread across the monitored wards, whereas CO₂ exceedance is concentrated in a smaller number of high-risk spaces, particularly neonatal and paediatric wards. Model results show that the TFT meaningfully outperforms XGBoost for temperature and RH forecasting, reducing error by approximately 32% and 35–42%, respectively, across MAE, RMSE, and MAPE, while CO₂ forecasting remains more challenging and is better served by the XGBoost. The proposed framework provides a basis for targeted ventilation management and demonstrates the feasibility of low-cost, health-oriented indoor climate forecasting in low-resource tropical hospitals.

Authors: Tekler, ZD; Roberts, B; Low, R; Cook, M

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: Air infiltration and Ventilation Centre
• Host title: Proceedings of the 46th AIVC Conference in Innovations in Smart Ventilation and IEQ for Resilient and Adaptive Buildings
• Publication date: 2026-10-01
• Acceptance date: 2026-08-05
• Event title: 46th AIVC Conference in Innovations in Smart Ventilation and IEQ for Resilient and Adaptive Buildings (AIVC 2026)
• Event location: Incheon, Republic of Korea
• Event website: https://aivc2026conference.org/
• Event start date: 2026-09-30
• Event end date: 2026-10-01
• Language: English
• Keywords: overheating; indoor environmental quality; natural ventilation; forecasting; machine learning; hospitals