Data-driven three-stage polymer intrinsic viscosity prediction model with long sequence time series data

Abstract

Intrinsic viscosity is a critical evaluation indicator for polymer quality during the aggregation process. Accurate prediction of intrinsic viscosity is essential for the online monitoring and control of polymer quality. The aggregation process is a typical industrial production mode influenced by production continuity and varying sensor spatial locations. The data from this production process exhibits characteristics such as parameter redundancy, long time sequences, and time delays. Existing methods struggle to accurately match data features, resulting in poor prediction accuracy. To address these issues, we propose a three-stage method for predicting polymer intrinsic viscosity. Firstly, we introduce a key feature extraction method based on the maximum information coefficient and an approximate Markov blanket model to address the redundancy in aggregation process data parameters. Secondly, we propose a time delay analysis method based on cross-correlation to estimate time delay relationships and match time delay features in the aggregation process. Finally, we develop an improved information prediction model for long-term sequence prediction, utilizing locally sensitive hash attention to replace ProbSparse self-attention. This ensures the prediction effect remains effective as the sequence length increases. The results demonstrate that our proposed method reduces feature redundancy, calculates delay times similar to simulation results, and offers significant advantages in prediction accuracy compared to other time series prediction models under different sequence lengths.