Supervised Long Short Term Memory Network for Quality Prediction in Industrial Polyethylene Process with Limited Samples

To improve the quality prediction performance in industrial polyethylene process with limited samples, a supervised long short term memory (SLSTM) network is applied and modified in this study. In the proposed method, the historical samples of other grade are used for model training, i.e. the structure and parameters of SLSTM network. Then the trained model is fine-tuned by the limited available samples of the target grade. So that only some weight parameters are further modified based on the limited samples. In this way, the qualities of key variables can be fast and accurately predicted by the proposed method. By comparison, the prediction results in the industrial polyethylene process are obviously improved by the proposed method, showing advantages than the existing methods.