Browse by author
Lookup NU author(s): Ngiap Koh, Dr Anurag SharmaORCiD, Dr Jianfang Xiao, Professor Cheng Chin
This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
Neural networks utilizing backpropagation are a powerful tool for solar irradiance forecasting, which is vital for climate research and energy market operations. This study addresses the challenge of modeling complex, nonlinear relationships between weather parameters by introducing an innovative adaptive B-spline activation function. The piecewise polynomial approach proposed in this work, integrated into the neural network framework and optimized using the Adaptive Moment Estimation (Adam) algorithm, allows for effective parameter tuning after each epoch, resulting in enhanced forecasting accuracy. Unlike traditional activation functions, which suffer from issues like the "dying ReLU" problem, the adaptive B-spline function provides smooth, flexible mappings with continuous gradients, allowing it to capture intricate data patterns effectively. This adaptability makes it particularly suitable for high-precision environmental applications. Validation using real-world datasets from Singapore and Hawaii shows that the adaptive B-spline significantly outperforms conventional activation functions, delivering up to a 10% improvement in forecasting accuracy for both training and testing datasets. Furthermore, its robustness across various neural network architectures demonstrates its adaptability and compatibility with backpropagation. This research highlights the potential of optimized B-spline activation functions to improve the accuracy and dependability of neural network-based forecasting models.
Author(s): Koh NT, Sharma A, Xiao J, Chin CS, Xing CJ, Woo WL
Publication type: Article
Publication status: Published
Journal: IEEE Access
Year: 2025
Volume: 13
Pages: 85156 - 85169
Print publication date: 13/05/2025
Online publication date: 13/05/2025
Acceptance date: 08/05/2025
Date deposited: 03/06/2025
ISSN (print): 2169-3536
ISSN (electronic): 2169-3536
Publisher: IEEE
URL: https://doi.org/10.1109/ACCESS.2025.3569525
DOI: 10.1109/ACCESS.2025.3569525
Altmetrics provided by Altmetric
