Core Concepts
The author proposes a Physics-Informed Neural Network (PINN) approach to accurately predict temperatures in reheating furnaces, addressing the challenge of limited real-world data availability.
Abstract
The content discusses the application of zone method-based physics-informed neural networks for predicting temperatures in energy-intensive reheating furnaces. By incorporating prior physical knowledge, the proposed PINN model aims to improve generalization capabilities. The study compares the performance of the PINN model against classical machine learning baselines and analyzes various aspects such as hidden layer configurations, batch sizes, activation functions, and individual regularization terms.
The foundation industries play a crucial role in providing materials for various industries. Reheating furnaces are highlighted as energy-intensive components within these industries. Accurate temperature prediction is essential for reducing energy consumption and achieving Net-Zero goals. Due to challenges in obtaining real data, a computational model based on Hottel's zone method is used to generate training data for neural networks.
The zone method mathematically models energy flow in different zones of a furnace, enabling accurate predictions. The proposed PINN model incorporates Energy-Balance regularizers to enhance generalization capabilities. Experimental results show that the PINN outperforms traditional machine learning methods and highlights the importance of considering physical constraints in predictive modeling.
Stats
RMSE tG (↓): 10.04
RMSE tS fur (↓): 7.95
RMSE tS obs (↓): 31.64
MAE tG (↓): 6.84
Quotes
"The classical Hottel’s zone method provides an elegant way to model high-temperature processes inside reheating furnaces."
"Our proposed PINN model improves generalization capabilities by incorporating prior physical knowledge using Energy-Balance regularizers."