A large water flood system consisting of pumps and a pipeline network suffers from internal constraints, as demonstrated by comparing measured data with analytical predictions. These constraints result in sub-optimum injection rates, translating to reduced oil production. Various modifications had been made to the system. However, imperfect data and other variations in system performance made it difficult to quantify performance improvements. Pressure and flow rate data were collected before changes were implemented. This was used to train and test a machine-learning model. The machine-learning model was then successfully used to compare measurements after changes and predicted values to quantify enhancements.

The results include comparing classification and regression algorithms in supervised learning using several standard testing strategies. Algorithms tested include linear regression, random forests, k-nearest neighbours (KNN), neural networks and adaptive boost (AdaBoosting). The results indicated that some algorithms could predict network performance with an accuracy of several orders of magnitude less than one per cent. The use of machine learning is significantly more practical and reliable than the alternative of using calibrating analytical models and the method has wide applicability to network performance measurement and, to a certain extent, forecasting if care is taken over extrapolation.

This technique has not been used before in this way and offers a quick and easy-to-implement method for accurately assessing pipeline network performance based on imperfect measured data. The proposed paper aims to demonstrate the value of using machine learning to quantify hard-to-determine otherwise performance improvements in large water flood systems as parts of valuable oil production assets.