Data-driven models of crude distillation units for production planning and for operations monitoring

This work develops best practices for building consistent data-driven CDU models for production planning and for monitoring. Prior knowledge-based transformations of raw data are shown to be essential to achieve high model accuracy. Data-driven SOM-ResNet and PCA-ResNet deep models, feedforward neural network (FNN), partial least-squares (PLS), and least absolute shrinkage and selection operator (LASSO) are studied. Instead of increasing the dimensionality of data (as in SOM-ResNet), it is better to decrease the dimensionality via PCA and employ deep learning image processing ResNet model. In addition, increasing the number of residual blocks in the PCA-ResNet model increases its accuracy, while adding more hidden layers to FNN does not. With appropriate data selection and transformation, when the number of data samples is less than 500, LASSO model is the best for planning and for monitoring. If the number of data samples exceeds 3,000, then PCA-ResNet model is the best for monitoring.