摘要
Effective steam management supports cost control and carbon abatement in industrial processes. However, steam monitoring in industrial records often exhibits mixed sampling intervals. The mismatch in time interval creates a limited-data problem that conventional energy models often struggle to handle. Therefore, a model-agnostic meta-learning framework integrated with an attention-based long short-term memory network is proposed for steam-consumption prediction under limited-data conditions. Meta-training on related high-frequency source units learns shared attention parameters and enables rapid adaptation to a low-frequency target unit without requiring synthetic data generation. The performance of steam consumption prediction is validated using a large-scale case study of the crude glycerin purification process. The results demonstrate that the attention-based long short-term memory model outperforms traditional models with the highest coefficient of determination value (R2) of 0.772. The incorporation of meta-learning further enhances the prediction performance of the model, with a decrease in the prediction error from 168.891 to 123.777 kg/h and an improvement in R2 of 0.847. Furthermore, the energy-saving analysis indicates the reduction in annual steam consumption and greenhouse gas emissions of 4372.304 (11.63% reduction) and 613.815 tons, respectively.
•Meta-learning model for multi-rate industrial steam prediction is proposed.•Shared-attention meta-learner adapts from high- to low-frequency data tasks.•No synthetic data required to achieve robust prediction under limited data.•26.7% error drop and 9.7% R2 gain over AM-LSTM without meta-learning.•Annual savings of 4372 tons of steam and 613.8 tons of CO2eq in glycerin refinery.