Online Learning Control with Echo State Networks of an Oil Production Platform

The design of a control algorithm is exceptionally hard when models are unavailable, the physics are varying in time, or structural uncertainties are involved. One such case is an oil production platform in which reservoir conditions and the composition of the multiphase flow are not known precisely. Today, with streams of data being generated from sensors, black-box adaptive control emerged as an alternative to the control of such systems. In this work, an online adaptive controller based on Echo State Networks (ESNs) is employed in diverse scenarios of controlling an oil production platform. The ESN learns an inverse model of the plant from which a control law is derived to attain set-point tracking of a simulated model. The analysis considers high steady-state gains, potentially unstable conditions, and a multi-variate control structure. All in all, this work contributes to the literature by demonstrating that online-learning control can be effective in highly complex dynamic systems (oil production platforms) devoid of suitable models, and with multiple inputs and outputs.