Cloud point prediction model for polyvinyl alcohol production plants considering process dynamics

• We construct soft sensors that use machine learning to predict product properties. • Transfer learning was used to combine two datasets of product properties. • The time delay of the process variables was taken into account. • To evaluate their effectiveness, a model was constructed using actual plant data. • The proposed method allowed the construction of a highly accurate prediction model. In polyvinyl alcohol (PVA) production plants, the cloud point is controlled as one of the product properties. The cloud point is the temperature at which the water solubility rapidly decreases because hydrogen bonds between PVA and water are broken. If the cloud point falls below a controlled value, the performance of the PVA as a product deteriorates, and thus it is necessary to monitor the cloud point in real time and ensure that it does not deviate from the controlled value. However, the cloud point cannot be analysed frequently in practical operations. In this study, a soft sensor that continuously predicts the cloud point from process variables that can be measured in real time was constructed using machine learning. Furthermore, the prediction accuracy of the model was improved by (1) increasing the number of samples using a data set of similar quality to the cloud point, considering that the number of samples near the control value of the cloud point is small, and (2) selecting the optimum process variable and time-delay range, considering the time delay of the process variable to accommodate the dynamic behaviour of the plant. Analysis was conducted using data measured from an actual PVA manufacturing plant. When comparing the conventional method with the proposed method, the coefficient of determination, used as an indicator to evaluate model performance, improved from 0.288 to 0.729. Furthermore, the root mean squared error decreased from 1.52 to 0.937. These results demonstrate that the proposed method contributes to improved prediction accuracy.