Differentiation Tool Efficiency Comparison for Nonlinear Model Predictive Control Applied to Oil Gathering Systems

Abstract

This paper presents a comparison of gradient computation techniques required to solve a single-shooting formulation of nonlinear model predictive control (NMPC) problems. An oil production system with network structure is considered as test instance. The structure of the network is exploited to improve computational efficiency. Exact gradient sensitivity calculation methods (forward and adjoint) are compared along with the finite difference approximation. Forward and Reverse automatic differentiation for calculating Jacobians are also compared along with the finite difference approximation counterpart. Since there is a trade off involving accuracy and speed when calculating these gradients, the best combination of tools is case dependent and it is determined by the analyses of performance indexes arising when solving specific NMPC problems. A hybrid approach combining finite difference Jacobian calculations with adjoint sensitivity calculations gave the best performance for our test problems.