A Rajaperumal T, Chinnappan Christopher Columbus
School of Electrical Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu, India.
School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, Tamil Nadu, India.
Sci Rep. 2025 Aug 8;15(1):29102. doi: 10.1038/s41598-025-13306-7.
This study presents a comprehensive hybrid forecasting framework that synergizes machine learning algorithms, MATLAB Simulink-based physical modeling, and Physics-Informed Neural Networks (PINNs) to advance wind power prediction accuracy for a 10 kW Permanent Magnet Synchronous Generator (PMSG)-based Wind Energy Conversion System (WECS). Using a complete annual dataset of 8,760 hourly wind speed observations from the MERRA-2 platform, ten machine learning algorithms were systematically evaluated, including Random Forest, XGBoost, and an advanced Stacking ensemble model. The Stacking ensemble demonstrated superior performance, achieving an exceptional R of 0.998 and RMSE of 0.11, significantly outperforming individual algorithms. A detailed MATLAB Simulink model was developed to replicate turbine behaviour under identical wind conditions, physically, providing robust validation for ML predictions. The Simulink model achieved satisfactory performance under nominal wind conditions but exhibited computational constraints during extreme wind scenarios, leading to compromised output reliability. To bridge the gap between pure data-driven learning and physical realism, a Physics-Informed Neural Network was subsequently integrated to combine data-driven learning with physical constraints, using both observational data and physics-based synthetic datasets. Comparative analysis revealed that ML models deliver superior speed and accuracy for operational forecasting, while the PINN framework maintains physical consistency with competitive predictive performance. The framework's practical applicability was demonstrated through a 2026 case study for southern Tamil Nadu, which incorporated projected environmental changes, including a 0.6% annual decline in wind speed. This real-world validation showcased the framework's adaptability to evolving climatic conditions and long-term forecasting capabilities. This integrated methodology provides a robust foundation for enhancing wind power integration into modern energy systems, while maintaining both computational accuracy and physical interpretability, thereby supporting sustainable energy transition goals.
本研究提出了一个全面的混合预测框架，该框架将机器学习算法、基于MATLAB Simulink的物理建模以及物理信息神经网络（PINNs）相结合，以提高基于10千瓦永磁同步发电机（PMSG）的风能转换系统（WECS）的风电预测精度。利用来自MERRA-2平台的8760个每小时风速观测值的完整年度数据集，系统评估了十种机器学习算法，包括随机森林、XGBoost和先进的堆叠集成模型。堆叠集成模型表现出卓越的性能，相关系数R达到了0.998，均方根误差RMSE为0.11，显著优于单个算法。开发了一个详细的MATLAB Simulink模型，以在物理上复制相同风况下的涡轮机行为，为机器学习预测提供有力验证。Simulink模型在标称风况下取得了令人满意的性能，但在极端风况下表现出计算限制，导致输出可靠性受损。为了弥合纯数据驱动学习与物理现实之间的差距，随后集成了物理信息神经网络，将数据驱动学习与物理约束相结合，使用观测数据和基于物理的合成数据集。对比分析表明，机器学习模型在运行预测方面具有更高的速度和准确性，而物理信息神经网络框架在保持物理一致性的同时具有具有竞争力的预测性能。通过对印度泰米尔纳德邦南部的一个2026年案例研究，展示了该框架的实际适用性，该研究纳入了预计的环境变化，包括风速每年下降0.6%。这种实际验证展示了该框架对不断变化的气候条件的适应性和长期预测能力。这种综合方法为增强风电融入现代能源系统提供了坚实的基础，同时保持了计算精度和物理可解释性，从而支持可持续能源转型目标。
