Utilizing the Kolmogorov-Arnold Networks for chiller energy consumption prediction in commercial building

Abstract

Accurate prediction of chiller energy consumption is essential for optimizing energy usage and reducing operational costs in commercial buildings. Traditional predictive methods often struggle to capture the complex, nonlinear relationships inherent in energy consumption data. This study proposes the use of Kolmogorov-Arnold Networks (KAN) to address this challenge, leveraging their ability to model intricate nonlinear dynamics with high precision. The study introduces KAN as a novel application for real-world chiller energy prediction, using actual data obtained from a commercial building. The methodology involves comparing KAN's performance with Artificial Neural Networks (NN) and a hybrid metaheuristic algorithm combined with deep learning, namely the Teaching-Learning-Based Optimization with Deep Learning (TLBO-DL). The results show that KAN achieves an R2 value of 0.9465 and an RMSE of 6.1023, outperforming NN (R2: 0.9281, RMSE: 6.7709) and TLBO-DL (R2: 0.9366, RMSE: 6.2892). The novelty of this research lies in the innovative application of KAN to chiller energy consumption prediction, coupled with advanced parameter tuning and improved computational efficiency. This study not only demonstrates the superior accuracy of KAN but also contributes to the field by showcasing its practical utility and effectiveness in energy management systems.