Model-free control of compressors in an air network with unknown future air demand

Jeroen Willems, Denis Steckelmacher, Bruno Depraetere, Ann Nowe

Abstract:

Optimal control of complex systems often requires access to an exact or almost-exact model,
and information about the (future) external stimuli applied to the system (load, demand, ...).
This is particularly true in the case of air networks, in which compressors have to fill an air
tank, usually proportionally small compared to the production of the compressors and the
average downstream demand of air. The demand of air therefore largely impacts the pressure
in the tank, and the compressors have to react quickly to changes in demand. In this paper,
we propose a method based on Reinforcement Learning to produce a high-quality controller
for 3 compressors connected to the same air network. The Reinforcement Learning agent
does not assume any model (so the compressors, tubes, losses and demand do not have to be
modeled) and does not observe the future demand of air, or an approximation of it. Still, the
learned controller performs comparably to a highly-tuned Model Predictive Controller, and
largely outperforms MPC when even a small error exists in the predicted future demand.
This demonstrates that Reinforcement Learning allows to produce high-quality controllers
in challenging industrial contexts.