Determination the Levels of Thief Zones Based on Machine Learning

Abstract The technique of interwell tracer testing is considered as one of the most effective method to identify the thief zone (TZ) in reservoirs. However, in heavy oil reservoirs, tracer breakthrough curves are mostly parabolic and unimodal, thus resulting in slight differences between curves. It is inefficient and inaccurate to identify different types of curves with traditional methods applied to characterize the levels of TZs. In this paper, convolutional neural network (CNN) is applied to construct a classification model for the automatic identification of the levers of TZs. According to the TZs criteria specified on the field, the analytical tracer transport model was applied to generate 3000 curves as the sample, which can meet the requirements of model training accuracy. In the meantime, One-hot encoding, Xavier initialization, Adam optimizer, and mini-batch normalization were used to construct the model, and the key parameters are optimized to improve the performance of the model. The results show that the appropriate activation function is ReLU and the optimal dropout rate is 0.5. Moreover, the construction of CNN with discrete data points (DDP-CNN) as input contributed to a further improvement of classification accuracy of tracer curves. The accuracy of DDP-CNN in training set is 0.96, which is 14% and 23% higher than random forest (RF) and k-means, respectively. In practical applications, DDP-CNN proves capable to correctly classify 88 of the 100 curves.