Neural network assisted temperature and humidity field simulation method in tobacco curing process
Weimin Guo, Qiang Xu, Aiguo Wang, Guangqing Chen, Shuoye Zhou, Jianwei Wang, Zuxiao Chen, Jianjun Liu, Ding Yan, Junying Li, Xianjie Cai, Xuchu Chen, Yanling Zhang
Abstract
During the tobacco curing process, uneven temperature and humidity distribution within the curing chamber significantly affects tobacco quality. Effective management of these environmental factors is essential for producing high-quality tobacco. Traditional simulations using COMSOL Multiphysics struggle to accurately predict the dynamic conditions in curing barns. This study presents a novel hybrid model that combines neural network-based predictions with COMSOL simulation outputs to achieve highly accurate forecasting of temperature and humidity fields. COMSOL models the curing chamber, providing gradient information on temperature and humidity distributions at any given time. The model leverages spatially discrete time-series data with a neural network algorithm to accurately predict future environmental conditions. To obtain continuous spatial distributions, the Poisson Blending algorithm integrates the neural network’s discrete predictions with COMSOL’s spatial gradients, resulting in a comprehensive and precise representation of the curing environment. The findings show significant improvements in prediction accuracy, with temperature field errors ranging between −1 ∘ C and 1 ∘ C, and humidity field errors between −0.75 g∕m 3 and 1 g∕m 3 , outperforming traditional COMSOL simulations. conpared with methods such as linear regression (LR), long short-term memory (LSTM), gated recurrent unit (GRU) and Transformer, this algorithm achieves smaller average error ranges. This method enables accurate prediction of temperature and humidity distribution patterns within the curing chamber, supporting rational classification and loading of tobacco leaves and facilitating intelligent, precise control during curing. Consequently, it reduces losses and enhances efficiency. Ongoing researches are essential to fully leverage the benefits of this innovative technology. • The gradient variations of temperature and humidity within the curing barn were obtained using COMSOL. • High-precision time series predictions at discrete spatial points were achieved through a neural network algorithm. • The results from COMSOL and the neural network were integrated using the Poisson blending algorithm. • Yielding high-accuracy spatiotemporal predictions of the temperature and humidity fields. • The average temperature error was within 1 °C, and the average absolute humidity error was within 1 g/m 3 .