Machine-Learning-Assisted Identification of Steam Channeling after Cyclic Steam Stimulation in Heavy-Oil Reservoirs

Abstract

Cyclic steam stimulation (CSS) is one efficient technology for enhancing heavy-oil recovery. However, after multiple cycles, steam channeling severely limits the thermal recovery because high-temperature steam preferentially breaks through to the producers. To solve the issues of steam breakthrough, it is essentially important and necessary to recognize steam channeling. In this work, a machine-learning-assisted identification model, based on a random-forest ensemble algorithm, is developed to predict the occurrence of steam channeling during steam huff-and-puff processes. The set of feature attributes is constructed based on the permeability ratio, steam quality, and steam-injection speed, which provides the reference for the construction of the training-sample set, steam-channeling reconstruction set, and prediction set. Based on the realistic data, the Pearson correlation coefficient is implemented to confirm the linear correlation among different characteristics; thus, the dimension reduction of the characteristic parameters is achieved. The random oversampling method is adopted to treat the unbalanced training-sample set. Our results show that this model can accurately describe the current state of steam channeling and predict steam propagation in the following cycles.