Optimization of Extreme Learning Machine using Barnacles Mating Optimizer for chiller cooling load prediction in commercial building

Accurate prediction of chiller cooling performance is crucial for enhancing energy efficiency and ensuring operational reliability in heating, ventilation, and air-conditioning (HVAC) systems. Conventional prediction approaches and basic machine learning models often face limitations, such as low generalization capability and reduced accuracy when dealing with complex and nonlinear cooling patterns. To address these challenges, this study proposes an Extreme Learning Machine (ELM) optimized with the Barnacle Mating Optimizer (BMO) for chiller cooling load prediction. The BMO algorithm was selected due to its strong capability to balance exploration and exploitation in the search process, inspired by the barnacle's natural mating behavior. Compared to many existing metaheuristic algorithms that often face issues such as premature convergence or limited local search performance, BMO introduces an adaptive mating mechanism that enhances convergence stability and maintains population diversity, making it theoretically and practically suitable for optimizing ELM parameters. The proposed model was evaluated using a benchmark dataset widely applied in chiller fault detection and diagnosis research, which records the operating parameters of a water-cooled chiller system installed in a commercial building located in a subtropical climate region. Model performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and symmetric mean absolute percentage error (sMAPE). The proposed BMO-ELM model was benchmarked against ELM optimized with the Dung Beetle Optimizer (DBO), Zebra Optimization Algorithm (ZOA), Genetic Algorithm (GA), as well as baseline single ELM,Linear Regression (LR) and Support Vector Machines (SVM) models. The experimental results show that BMO-ELM achieved superior performance in MAE (1.11) and sMAPE (8.20), and competitive results in RMSE (1.92) and R2 (0.92). This highlights its potential to enhance prediction accuracy and address model limitations, making it suitable for chiller cooling prediction.