Hottel Zone Physics-Constrained Networks for Industrial Furnaces

Abstract

This paper investigates a novel approach to improve the temperature profile prediction of furnaces in foundation industries, crucial for sustainable manufacturing. While existing methods like the Hottel Zone model are accurate, they lack real-time inference capabilities. Deep learning methods excel in speed and prediction but require careful generalization for real-world applications. We propose a regularization technique that leverages the Hottel Zone method to make deep neural networks physics-aware, improving prediction accuracy for furnace temperature profiles. Our approach demonstrates effectiveness on various neural network architectures, including Multi-Layer Perceptrons (MLP), Long Short-Term Memory (LSTM), Extended LSTM (xLSTM) and Kolmogorov-Arnold Networks (KANs). We also discussion the data generation involved.

Keywords

• Extended LSTM (xLSTM)
• furnace
• Hottel zone method
• Kolmogorov-arnold networks (KANs)
• long short-term memory (LSTM)
• multi-layer perceptrons (MLP)