A novel ensemble network based on <scp>CNN</scp> ‐ <scp>AM</scp> ‐ <scp>BiLSTM</scp> learner for temperature prediction of distillation columns

Abstract In recent years, complexity has significantly increased in chemical processes where a distillation column serves as a crucial unit. It is worthwhile to develop an accurate and reliable predictive model to maintain the steady operation condition of distillation column. Although data‐driven models that do not rely on any prior knowledge present a promising approach, they encounter challenges associated with nonlinearity and dynamic behaviour within process data. To tackle these challenges, a deep learning‐based combined distilled spatiotemporal attention ensemble network (CDSAEN) is proposed. The CDSAEN is constructed by sequentially integrating multiple base learners, which are iteratively distilled and generated with decreasing attention span lengths through the boosting method implemented by a specially designed attention extraction evaluation function. In a base learner, convolutional neural network (CNN), attention mechanism (AM), and bidirectional long short‐term memory (BiLSTM) are utilized to adaptively capture deep and intricate spatiotemporal features and establish a robust mapping relationship from inputs to output. Real‐world process data from a distillation system in a chemical plant is reconstructed as a time series dataset and is subsequently fed into CDSAEN for training to forecast the temperature of the distillation column apparatus in advance. The results exhibited effectiveness and reliability. Additionally, in comparison to six other data‐driven predictive approaches, the proposed method attained superior performance with mean absolute error (MAE) = 0.084, root mean squared error (RMSE) = 0.108, and R 2 = 0.974. This study can provide support for maintaining the stable operation of distillation columns in chemical processes.