Semi-Supervised Caustic Concentration Prediction in Alumina Production with Transformers

Accurate and real-time prediction of caustic concentration is essential for optimal control in alumina production. However, training effective models is often hampered by the scarcity of labeled assay data. This paper proposes a novel semi-supervised learning approach to address this challenge. We introduce a Transformer-based Auto-Encoder (AE) structure that uses both labeled and unlabeled data for robust caustic concentration prediction. The Transformer encoder extracts key features from various variables and time series data, enabling efficient low-dimensional representation. The decoder then performs a dual task of reconstructing the input data and predicting the caustic concentration. This structure combined with a robust regularization strategy enhances the model’s resilience to data noise and missing values. Consequently, the proposed method achieves accurate caustic concentration prediction using significantly less labeled data. This approach has the potential to significantly improve alumina production control by reducing reliance on expensive and time-consuming labeled data collection.