CO emission prediction in rotary dryers: a temporal fusion transformer approach

Abstract

Carbon monoxide (CO) emission control in rotary dryer systems is critical for preventing explosive incidents within the Electrostatic Precipitator (ESP), particularly when emission levels approach the threshold of 10,000 ppm. Conventional reactive control systems relying solely on automatic shutdown upon threshold exceedance are insufficient for proactive hazard mitigation. This study proposes the application of the Temporal Fusion Transformer (TFT) for multi-horizon prediction of CO gas emissions in a nickel drying rotary dryer system, representing the first such application in this industrial context. The dataset comprised 691,182 entries of operational parameters and laboratory results from a nickel smelter, preprocessed using the Savitzky-Golay filter to reduce signal noise while preserving critical data features. The TFT model was evaluated against five benchmark models: Generalized Additive Model (GAM), Neural Network Regression (NNR), Bagged Regression Tree (BRT), Linear Support Vector Machine (LSVM), and Long Short-Term Memory (LSTM), using RMSE, MAPE, and R-squared as the evaluation metrics. The TFT achieved superior performance with an RMSE of 2.42, MAPE of 1.58%, and R-squared of 0.999, substantially outperforming all competing models. Beyond its predictive accuracy, the TFT variable selection network provides interpretable insights into the operational parameters that most strongly influence CO emission levels, enabling data-driven decision making for process operators. These results demonstrate that TFT effectively transforms emission control from a reactive to a proactive, prediction-based paradigm, thereby minimizing production disruptions while enhancing process safety.