Air conditioning load prediction using the automated model mixture of experts framework

Accurately predicting the air conditioning (AC) load in commercial buildings is essential for improving energy efficiency and supporting low-carbon operational strategies. However, this task is challenging due to the non-linear and dynamic influence of multiple factors, such as building characteristics, environmental conditions, occupancy patterns, and temporal variations. To address this challenge, this study develops a long-term dataset from a commercial office building, integrating architectural, environmental, and operational attributes. Based on this dataset, an Automated Mixture-of-Experts (AMMoE) framework is proposed, which adaptively selects and combines predictive models in response to changing data characteristics. The framework continuously evaluates model performance, designates primary and supporting experts, and aggregates their outputs through adaptive weighting, enabling accurate representation of heterogeneous time-series building data. Experimental results show that AMMoE markedly outperforms individual models, achieving a Normalized Mean Bias Error (NMBE) of -0.38%, a Coefficient of Variation of Root Mean Square Error (CVRMSE) of 34.52%, and an R² of 0.96. This confirms its robustness in real-world operational settings. The main contribution lies in establishing a practical, adaptive forecasting approach that links ensemble modeling principles with engineering applications. This work demonstrates the value of adaptive expert integration in advancing building energy management and provides a basis for future cross-building generalization.