Accurate cooling load prediction in chiller systems is critical for optimizing energy efficiency in commercial buildings, where HVAC systems account for 50%–60% of total energy consumption. Traditional prediction methods fail to capture complex non-linear relationships, while conventional neural network training suffers from local optima issues. This study proposes a novel hybrid approach combining the Barnacles Mating Optimizer (BMO) with Artificial Neural Networks (BMO-NN) for enhanced cooling load prediction. The methodology employs a real-world dataset from commercial chiller operations, incorporating seventeen operational parameters, including temperature measurements, flow rates, and electrical parameters to predict cooling load. The BMO-NN model was evaluated against established hybrid metaheuristic-neural network including Particle Swarm Optimization (PSO-NN), Ant Colony Optimization (ACO-NN), Slime Mould Algorithm (SMA-NN), Reptile Search Algorithm (RSA-NN), and traditional ADAM optimization, using RMSE, MAE, and R2 metrics. SHAP (SHapley Additive exPlanations) analysis investigated feature importance patterns and model interpretability across algorithms. Results demonstrate BMO-NN’s superior performance, achieving RMSE of 2.8551, MAE of 1.8273, and R2 of 0.7440. The model exhibited exceptional stability with minimal performance variation (RMSE range of 0.16). The SHAP analysis indicated that the effectiveness of the BMO-NN model resulted from its ability to balance physically meaningful variables, particularly those related to electrical and thermal characteristics. These findings confirm that the integration of the BMO algorithm in NN training is effective for HVAC applications, offering building operators a reliable tool for proactive energy management and improved energy efficiency.