In recent years, various inferential models have been developed to predict product quality using first principles, linear regression or neural networks. The first principle models are able to survive process modifications (such as replacing trays, cleaning condensers or reboilers, etc.) quickly by changing equation coefficients but require sufficient process knowledge. Linear regression models and neural network models are empirical models which require less process knowledge (different from no prior process knowledge, for any inferential models to work well for the real plant, process knowledge is important) but require much more process data and has less ability to survive process modifications. Each methodology has its advantage and disadvantage and it's difficult to determine which inferential models take advantage over the others for general inferential modelling problems. Given proper understanding and insight, both the inferential models could give comparable results. For our case to predict a quality property in a refinery, it's not easy to build the first-principle model due to the complexity of the process while we have large volumes of process data, therefore empirical models rather than first-principle models are considered in this paper. Furthermore the process is basically a nonlinear process (such as vapour-liquid phase equilibrium) and operates at different feed and operation conditions (such as different crude feed), while a single linear regression model may not work well for a wide operation condition, multiple models or nonlinear transformation might be required for different operation conditions. Based on this consideration, neural network approach combined with process knowledge is selected in this paper.
In this paper, a data-driven soft sensor approach based on neural networks and process knowledge is presented, with the aim to predict a quality property of ASTM 90% distillation point temperature for a product in a refinery. This paper first gives a short description of the neural network approach using a Back Propagation (BP) network. Then it explains why this approach suits for quality prediction and why a single neural network may perform the function of multiple linear models which would be required for any linear regression approach to survive change of process operation conditions. This paper discusses the selection of process variables where the principal component analysis (PCA) could be helpful but the process knowledge is more important. This paper discusses the dead time presented in the analyzer data. This paper also presents other pre-processing steps and methods including removing missing data and bad data via process knowledge and statistical process control (SPC), filtering the data, as well as scaling/transforming the data. Then this paper presents the modelling approach where the training data set is divided into three parts: training, validation and test. When train error, test error and validation error are minimized, the neural network model is saved and becomes the model. The model is then verified by running it over a totally new set of data and the prediction output of the model matches the measurement data well.
This paper presents a neural-network-based soft sensor for a refinery application and some results have been achieved. The prediction output of the soft sensor matches the lab measurement data better than the online analyzer and it eliminates the dead time, thus it may be used in close-loop control to improve product quality and process operation. The results also show that this inferential model may deal with change of some operation conditions (such as different crude, etc.), but it may not be able to deal with process modifications. To survive significant process modifications, any inferential models would need to be re-calibrated.