Data driven fast real-time optimization with application to crude oil blending

Real-time optimization (RTO) uses the information from the current process to update the model and predict the optimal operational policy. There are two difficulties for RTO in fast processes: 1) RTO needs a model to describe the process; 2) Optimization solutions come from slow recursive calculations. In this paper, we propose a data-based approach to overcome these two drawbacks. First we use historical data and static neural networks to approximate the output of the process. Then, the recurrent neural networks are applied to solve the optimization problem. So the novel RTO does not need a model and is much faster than the classic one. Finally, we apply this novel method to the crude oil blending. The results of the simulation are provided to illustrate the successful application of the data driven fast RTO.