A Two-Stage Deep Learning Framework for Uncertainty-Aware Forecasting and Conditional Process Optimization Toward Sustainable Smart Manufacturing

The rise of Industry 4.0 has made robust forecasting and process control central to achieving sustainable smart manufacturing. While deep learning improves prediction accuracy, most models lack uncertainty quantification and process control—both critical for minimizing waste and optimizing energy use. Generative models are promising but underused in real-world manufacturing due to data and integration challenges. To address this, we propose a unified two-stage deep learning framework for real-time simulation and adaptive process control in dynamic manufacturing environments. The first stage leverages DeepAR and Monte Carlo Dropout to produce both accurate point predictions and calibrated uncertainty intervals. The second stage introduces a novel LSTM-based Conditional Variational Autoencoder (LSTM-CVAE) that reconstructs temporally coherent multivariate input sequences conditioned on target specifications. Then a two-stage filtering mechanism ensures the plausibility of generated inputs and their predictive alignment with operational goals. Experiments on two real-time industrial datasets (sugar and bioprocessing) and a public benchmark demonstrate average improvements of up to 35.7%, 21.0%, 24.6% in Balanced MAE, and 21.3%, 9.8%, 25.2% in Comprehensive Uncertainty Score metrics. The framework remains robust under batch-wise shuffling, supporting real-world deployment. Overall, it enables interpretable, forecast-aware, target-driven process control that improves resource efficiency and product cost—advancing sustainable manufacturing.