Digital twin-aided transfer learning for energy efficiency optimization of thermal spray dryers: Leveraging shared drying characteristics across chemicals with limited data

Abstract

Efficient energy management is crucial for spray-drying units as it can substantially improve product yield, reduce operating costs, and enhance energy utilization. However, due to limited data problems, the monitoring performance of the energy efficiency of a model is inefficient and unreliable, making it difficult to adjust operating conditions and hindering effective utility management. Therefore, this study proposes a long short-term memory-based transfer learning model with shared source-target characteristics for enhancing energy efficiency trackability under limited efficiency labels. Utilizing a long short-term memory structure improves the capability of capturing the process dynamic behavior. Synchronously, the digital twin-aided transfer learning concept supports the model by leveraging the parameters learned from the simulated source domain to assist the performance of the model in a limited data domain with different chemicals. The reliability and accuracy of the model are verified by a real industrial case study involving the detergent powder drying process. Results show that the model testing achieved an r-squared value of 0.938, outperforming conventional techniques by boosting the performance of the network up to 14.53 % and reducing surplus energy on demand and supply by 50.05 % and 81.27 %, respectively. The proposed model reveals the interconnection between source and target accuracy and provides a reliable learning process of the target domain observed based on the distribution of the testing performance. Notably, the model deployment indicates a considerable decrease of 16.63 % in natural gas consumption, leading to an enhancement of 11.92 % in evaporation efficiency and the prevention of 483 tonnes of carbon emissions annually.