Model-free H tracking control for de-oiling hydrocyclone systems via off-policy reinforcement learning

Shaobao Li, Petar Durdevic, Zhenyu Yang*

*Corresponding author for this work

Research output: Contribution to journalJournal articleResearchpeer-review

11 Citations (Scopus)


In offshore Oil and Gas production, it is important to ensure the quality of discharge water meets government regulations; the de-oiling hydrocyclone system is important for this kind of water treatment. Hydrocyclone's de-oiling performance is very sensitive to the inflow variation, which is often introduced by the upstream three-phase separator's operation. Thereby, the coordination control of both the separator and hydrocyclone turns to be crucial. Many model-based control methods can be employed for handling this coordination if mathematical models of the considered systems are developed. However, to develop models of these interacted systems is far more than trivial, with respect to the situation that each offshore de-oiling installation at different fields is a kind of tailor-made solution. Thereby, instead of using conventional model-based control design methods, this paper investigates the reinforcement-learning-based H control method for synthesizing an advanced model-free control solution for the coordination control of the separator's (water) level and hydrocyclone's Pressure-Drop-Ratio (PDR). The H tracking control is formulated as a 2-player zero-sum differential game and derived by employing an off-policy reinforcement learning algorithm based on state feedback and output feedback, respectively. Stability and optimality of the proposed solution are analyzed. Finally, simulation studies demonstrate the effectiveness of the proposed solution.

Original languageEnglish
Article number109862
Number of pages12
Publication statusPublished - Aug 2021


  • De-oiling hydrocyclone system
  • H control
  • Output feedback control
  • Reinforcement learning
  • Three-phase separator
  • Zero-sum game


Dive into the research topics of 'Model-free H tracking control for de-oiling hydrocyclone systems via off-policy reinforcement learning'. Together they form a unique fingerprint.

Cite this