A Hierarchical Ensemble Model With Novel Channel–Spatial Attention Module for Multistep Industrial Temperature Prediction

Temperature prediction is crucial for improving the reliability and homogeneity of industrial ethanol precipitation. Due to the limitations of complex measurement environment and the high viscosity of solution, it is difficult to accurately predict the temperatures of ethanol precipitation tank. To address this problem, a transformer-based ensemble learning method is proposed to multi-step predict the temperature using multivariate process parameters. Firstly, a novel transformer-based model is presented to predict the temperature of ethanol precipitation tank. Secondly, considering the information sparsity of valve-type signals, a multi-scale feature extraction module is introduced by designing global average pooling network blocks with multi pool sizes to extract multi-scale temporal features from valve-type signals. Thirdly, a novel channel-spatial fused attention module is proposed to capture both global information and local feature. In addition, a novel hierarchical AdaBoost framework is introduced to further boost the performance of the proposed model. Extensive real-world experiments validate that our proposed model outperforms state-of-the-art models in terms of accuracy and robustness. In 10-minute industrial temperature prediction, our model achieves an average reduction of 28.03% in MAE and 34.85% in MSE, compared to existing studies.