(749d) Efficient Real Time Optimization Using Approximate Dynamic Programming

In this work, an efficient real time optimization framework is developed based on the approximate dynamic programming (ADP) with less computational demand and high robustness to the uncertainties.

Nowadays, due to serious competitions in the market and high price of energy, how to optimize the plant-wide operations to enhance the profitability is of course important to the entire chemical industry. The real time optimization (RTO), as a module to coordinate the operating conditions (setpoint) of different units, thus becomes more valuable in the process system engineering. The steady state RTO, even though requires less computational efforts, totally ignores the dynamics of the process and thus may lead to infeasible operations [1]. The dynamic RTO, even though takes dynamic model into account, may not be robust to the uncertainties and is computationally expensive because of the multistep prediction. Thus, it is necessary to consider some new methodologies to overcome all drawbacks mentioned above. The approximate dynamic programming (ADP) was introduced into process system engineering community for a decade [2]. It relies on the principle of optimality [3] to reduce online computational demand and makes use of simulation/historical data to find profitable operational patterns for plant-wide system.

We propose an ADP-RTO to reduce online computational work and achieve better economic objective values. First, through data analysis, the information of process states, setpoint and production is extracted from operational/historical data. Second, a surrogate model and approximate value function are constructed to characterize the transient and steady performances of a plant under certain control schemes. Third, in the online implementation, a small-scale mixed-integer linear program (MILP) is solved at each time instant to maximize the one-step ahead value function and then decide whether the setpoint should be changed such that the economic objective along the prediction horizon is optimized.

Several features of proposed ADP-RTO are shown in order. First, even though this ADP-RTO is based on the historical data, the real time measurement also can be used to continuously update the value function and surrogate model in order to reduce the impact of uncertainties. Second, this ADP-RTO can be integrated with any types of controller, not limited to the model predictive control (MPC). Third, this ADP-RTO avoid identifying nonlinear dynamic representation for the entire plant, because such modeling process can be time consuming, costly and very sensitive to the disturbances. Fourth, due to its less computational demand, the ADP-RTO can be triggered more frequently to optimize the dynamic behavior of the plant.

Reference

[1] Tosukhowong, T., Lee, J. M., Lee, J. H., Lu J. (2004). An introduction to a dynamic plant-wide optimization strategy for an integrated plant. Computers and Chemical Engineering, 29, 199-208.

[2] Lee, J. H., & Lee, J. M. (2006). Approximate dynamic programming based approach to process control and scheduling. Computers and Chemical Engineering, 30, 1603-1618.

[3] Bellman, R. E. (1957). Dynamic programming. Princeton University Press.