Decomposing the optimization of a gas lifted oil well network

Low pressure oil wells can be made to flow by the injection of gas into the well. The resulting flow from the well increases discontinuously with the amount of gas injected. A common way to model this is by introducing integer variables which results in a MINLP formulation. Such a problem can be solved efficiently by a specialised Branch & Bound code. However for more than about 200 wells a complete search of the Branch & Bound tree becomes intractable. In this report we show that problems with a certain structure can be solved by different decomposition methods, one of which uses a bundle method to solve the resulting master problem. Improvements of the solution times up to a factor of 100 were observed on the largest problems which could be solved without decomposition, and problems with 1000 wells were solved by the decomposition method. We also show that there is substantial scope for the parallelization of such a method.