Due to the significant increase in cooling demand, accurately predicting cooling energy consumption in buildings has become essential. Machine learning (ML) techniques have been widely used for this purpose, with ensemble methods, which combine the outputs of multiple individual models, showing strong performance in handling its complexity. The primary objective of this study is to provide a systematic analysis of various ensemble aggregation strategies, including some novel combinations. Both homogeneous and heterogeneous ensembles are analyzed and tested using the same real-world measured dataset. Random Forest (RF) and Gradient Boosting Machine (GBM), which represent homogeneous ensembles that utilize bagging and boosting techniques, respectively, have proven to outperform individual models. Four prominent base models, namely Feedforward Neural Network (FFNN), Decision Tree (DT), Support Vector Regression (SVR), and Gaussian Process Regression (GPR), are utilized to construct eight new heterogeneous ensembles. These base models are combined via weighted averaging, with various strategies for assigning weights systematically tested. The stacking/ blending approach is also evaluated using all four base models along with Multiple Linear Regression (MLR) as meta learners. Results show that weighted averaging enhances prediction accuracy, with the best results obtained by solving an optimization problem that minimizes the mean squared error (MSE). Additional improvements are seen when an ML model acts as the aggregator, with the highest performance achieved using an FFNN as the meta-learner. Despite its simplicity and low computational demands, MLR proves to be an effective meta-model, providing a solution that requires less extensive hyperparameter tuning compared to other nonlinear machine learning models.
Jovanović et al. (Wed,) studied this question.