Model updating in the presence of stochastic feedstock disturbances for real-time optimization of blending operations

Real-time optimization based control systems are used for blending operations in many industries. Typically, in blending of automotive gasoline, these systems use linear programming with bias update of blending models. This paper considers the performance of such systems in terms of profit level and shows that the blender control performance can be increased very significantly by eliminating or reducing plant/model mismatch. Further, a new control approach that optimizes over the entire blend, rather than at a point in time, is presented and the benefits of adopting this new approach are illustrated. The paper concludes with a brief discussion of observability issues which must be considered to ensure the success of the proposed control scheme.