Improved secretary birds optimized weighted ensemble for energy use prediction

• Metaheuristic-optimized heterogeneous weighted ensemble integrating FFNN, DT, SVR, and GPR for building energy prediction • Joint optimization of 16 parameters in total, including base model hyperparameters and ensemble weights, within a unified metaheuristic framework • Development of the ISBOA algorithm by enhancing SBOA with chaotic mapping and dynamic opposition-based learning (DOBL) • Sensitivity analysis of DOBL parameters ( J r and w ) to identify robust operating regions • Strong and consistent performance validated on both benchmark simulation datasets for heating and cooling loads and real-world cooling energy data • Outperforms individual learners, classical ensembles, and previously reported benchmark models. With the increasing demand for energy and its associated environmental impact, accurate prediction of building energy use has become essential. Machine learning approaches have shown strong potential for addressing this complex task, with ensemble methods offering improved predictive capability over individual models. This study proposes a novel heterogeneous weighted ensemble framework, ISBOA-ENS, designed to enhance prediction accuracy through advanced metaheuristic optimization. The ensemble integrates four distinct base learners: Feedforward Neural Network (FFNN), Decision Tree (DT), Support Vector Regression (SVR), and Gaussian Process Regression (GPR). To simultaneously optimize the hyperparameters of the base models and the ensemble weights, a new metaheuristic algorithm, the Improved Secretary Bird Optimization Algorithm (ISBOA), is developed. ISBOA extends the Secretary Bird Optimization Algorithm (SBOA) by incorporating Dynamic Opposition-based Learning (DOBL) and Chaotic Mapping (CM), enabling more effective exploration of the high-dimensional search space comprising 16 parameters in total. The proposed approach is evaluated using three datasets, including a widely used simulation-based benchmark dataset for heating and cooling loads and a real-world dataset containing measured cooling energy consumption and associated input variables. Across all test scenarios, ISBOA-ENS demonstrates consistently strong performance, outperforming individual base learners, the SBOA-optimized ensemble, classical ensemble methods such as Random Forest (RF) and Gradient Boosting Machine (GBM), as well as a constrained least-squares weighted ensemble. Furthermore, the proposed model achieves better predictive accuracy than all previously reported models on the benchmark simulation dataset.