The aim of this paper is to enhance oil production from gas-lifted oil wells by solving Mixed-Integer Linear Programs (MILPs). These programs need to be repeatedly solved as the parameters of the wells change. Instead of using expensive exact methods or approximate methods, the paper proposes a heuristic solution based on deep learning models. These models are trained to provide values for the integer variables by considering different well parameters. By fixing the integer variables early on, the original problem is simplified to a linear program (LP). The paper presents two approaches for developing the learning-based heuristic: a supervised learning approach that requires optimal integer values for training instances, and a weakly-supervised learning approach that only needs solutions for early-fixed linear problems with randomly assigned integer variables. The results demonstrate a significant reduction in runtime of 71.11%. Additionally, the weakly-supervised learning model produces valuable early fixing values, even without knowledge of the optimal values during training.
Early Fixing for Optimization of Gas-lifted Oil Production: A Comparison of Supervised and Weakly-supervised Approaches Using Deep Learning (arXiv:2309.00197v1 [cs.LG])
