An Ensemble Multi-Task Learning Model for Predictive Performance Evaluation of Air Handling Units in HVAC Systems

With urbanisation resulting in increased demand for indoor comfort, HVAC (heating, ventilation, and air-conditioning) systems, particularly air handling units (AHUs), are essential for indoor climate control. The advent of big data and artificial intelligence (AI) has opened new avenues for enhanced safety and reliability in HVAC operations. Hence, this study focused on the predictive performance evaluation of AHUs, which is receiving less attention compared to its fault detection and optimal control issues. Utilising real-time operational data from the Oak National Laboratory, the proposed model employs multi-task learning (MTL) to refine prediction accuracy for AHU return air properties, including temperature, moisture content, and power consumption. This is achieved without allowing any single task to dominate others during the training phase. Moreover, the model introduces an ensemble approach that synergises the capabilities of the different MTL algorithms using a boosting technique via a gradient boosting regression tree (GBRT). The study evaluated four MTL weighting strategies: average loss (MTL), geometric loss (MTL-GL), dynamic weighting (MTL-DW), and uncertainty weighting (MTL-UW). The baseline MTL achieved the highest accuracy for environmental variables, with an MSE of 0.0022, MAE of 0.0391, and an R2 score of 0.98 for return air temperature. In contrast, MTL-UW performed best for RTU power prediction, attaining an MSE of 6702, MAE of 45.6620, and an R2 of 0.88. In this light, we developed an ensemble multi-task learning (E-MTL) framework. The novel strategy significantly surpassed conventional data-driven baselines, as the E-MTL models consistently outperformed individual MTL variants, achieving the lowest RMSE and MAE across all target variables. This study bridges advanced artificial intelligence and environmental stewardship by mapping precise multi-task predictions directly onto building sustainability metrics. The paper culminates by showcasing the significant role of the proposed model as a metric for AHU performance evaluation and its contribution to smart decision-making in a real-world context. By enabling actionable operational foresight, the proposed model serves as an important framework for lowering HVAC operational carbon footprints, minimising net energy demands, and advancing sustainability goals. Essentially, the developed model is poised to facilitate optimal decision-making regarding HVAC components and foster proactive strategies to ensure consistent operation and extend the lifespan of HVAC systems.