Artificial intelligence-assisted performance prediction and operating-point identification for passive heat recovery ventilation systems

Heat Recovery Ventilation (HRV) systems reduce building energy demand by thermally coupling exhaust and fresh outdoor air streams through a plate heat exchanger core, without auxiliary heating. This study presents a machine-learning (ML) framework for predicting and optimizing the passive performance of the Argemsan K-203 cross-flow plate heat exchanger under variable airflow conditions. Outdoor fresh air at 7.8oC/81%RH was drawn through the HRV core and supplied to the occupied space, while indoor exhaust air at 21.1oC /55%RH was simultaneously discharged outside through the same core-the built-in 1500W electrical heater was intentionally deactivated throughout. Experiments were conducted over 60 minutes at 13 measurement states spanning u_t∈{2, 4, 6} m/s and u_e∈{2, 4, 6, 7.2} m/s. Three regression models - Random Forest (RF), Gradient Boosting (GB), and Ridge Regression-were evaluated via Leave-One-Out Cross Validation (LOO-CV). RF achieved R2 =0.931, MAE=0.261oC for fresh-air outlet temperature (T4); Gradient Boosting achieved R2=0.880 for sensible efficiency (ηT). Sensible efficiency ranged from 48.2% to 78.8% (mean 58.9%) and a humidity-ratio–based latent effectiveness of 35-59% (mean 44%). Recovered heat power spanned 245-713W, yielding 236-685 kWh/year of annual energy savings and 48-138 (natural gas) or 100-290 (grid) kgCO2/year avoidance per unit. Response-surface analysis identifies u_t=2 m/s with u_e ≥6 m/s as the optimal operating band.