Lessons Learned from Solving Data-Driven Optimal Control Problems as MILPs
October 2024
The talk was given in French.
An industry talk at a GDR RO&D (CNRS Operations Research) industrial day, in October 2024. A feedback session on turning real optimal control problems into mixed-integer linear programs at Purecontrol, when the system dynamics come from data rather than clean analytical models.
It works through the modelling choices that rarely make it into papers: what to make a decision variable, what to keep as a constraint, how to reach a MILP a solver can handle in production time. The biogas upgrading case is the running example. It is the formulation know-how behind the ROADEF 2024 application, where the same problem (MPC over a MILP) cut the upgrading unit’s operating cost by about 9%. The talk also presents pureSolve, the in-house library we use to build and deploy these models.
The tooling is not specific to energy. The same optimal-control formulation underlies my wastewater aeration control and the rest of the industrial control work across water and energy.
Slides are not shared publicly for this talk.
Joint work with François Gauthier-Clerc (Purecontrol).