Prediction of f-CaO content in cement clinker using a GRU-based deep learning model with masked-attention mechanism for incomplete DCS data

Efficient real-time prediction of free lime (f-CaO) content is critical for cement clinker quality control. However, process parameters from the distributed control system (DCS) often suffer from significant missing values, while time delays, nonlinear relationships, and coupling effects further hinder prediction accuracy. To overcome these challenges, this study presents a GRU-based f-CaO prediction model. Sixteen key parameters were first selected from 180 DCS-collected variables, then reconstructed and temporally aligned through the integration of empirical formulas, expert knowledge, and the time window principle. An encoder module equipped with a masked-attention mechanism processes missing values and extracts essential features, while the GRU framework predicts the f-CaO content. To mitigate data imbalance issues, a weighted loss function was applied during training, achieving 16.7%, 25.0%, and 17.2% reductions in MAE, MSE, and RMSE, respectively. The model achieves both high accuracy and low latency, meeting the needs of real-time and stable cement production.