Comparison of product quality estimation of propylene polymerization in loop reactors using artificial neural network models

One of the major challenges in polymerization industry is the lack of online instruments to measure polymer end-used properties such as xylene soluble, particle size distribution and melt flow index (MFI). As an alternative to the online instruments and conventional laboratory tests, these properties can be estimated using model based-soft sensor. This paper presents models for soft sensors to measure MFI in industrial polypropylene loop reactors using artificial neural network (ANN) model, serial hybrid neural network (HNN) model and stacked neural network (SNN) model. All models were developed and simulated in MATLAB. The simulation results of the MFI based on the ANN, HNN, and SNN models were compared and analyzed. The MFI was divided into three grades, which are A (10-12g/10 min), B (12-14g/10 min) and C (14-16 g/10 min). The HNN model is the best model in predicting all range of MFI with the lowest root mean square error (RMSE) value, 0.010848, followed by ANN model (RMSE=0.019366) and SNN model (RMSE=0.059132). The SNN model is the best model when tested with each grade of the MFI. It has shown lowest RMSE value for each type of MFI (0.012072 for MFI A, 0.017527 for MFI B and 0.015287 for MFI C), compared to HNN model (0.014916 for MFI A, 0.041402 for MFI B and 0.046437 for MFI C) and ANN model (0.015156 for MFI A, 0.076682 for MFI B, and 0.037862 for MFI C)