Prediction of Multilateral Inflow Control Valve Flow Performance Using Machine Learning

Abstract

Summary Intelligent multilateral well completions provide downhole flow rate, pressure, and temperature measurements at multiple well segments, which allows for a continuous spatiotemporal data stream. Such an extensive data input poses a challenging task to decide on the optimal strategy of manipulating the inflow control valve (ICV) settings over time for best performance. In this study, we investigate using machine learning to analyze and predict well performance given different ICV settings to ultimately maximize the well output. A commercial reservoir simulator was used to generate two synthetic reservoir models: homogeneous (Case A) and heterogeneous (Case B). These synthetic data were used to train, validate, and test machine learning models. The reservoir cases were generated on the basis of a segmented, trilateral producer completed with three ICV devices installed at tie-in segments. The data used were measurements of wellhead and downhole flow rates across ICV segments over a period of 4,000 days. A total of 1,330 experiments were conducted with an 8-day timestep, generating a total of 667,660 sample data points for each of Case A and Case B. Fully connected neural networks were used to fit the data, while model generalizability was enhanced using regularization techniques, namely L2 regularization and early stopping. Both random sampling and Latin hypercube sampling (LHS) methods were evaluated in constructing the training, validation, and testing splits. Trained with different sample sizes drawn from the 1,330 simulated data histories for the two reservoir models, the proposed neural network showed excellent results. Given only 10 simulated choices of ICV settings for training, the network proved capable of predicting oil/water production profiles at surface for both homogeneous and heterogeneous reservoir models with greater than 0.95 coefficient of determination (R2) when evaluated at unseen, test ICV settings. Extending the problem to downhole flow performance prediction, approximately 40 training simulated settings were necessary to achieve 0.95 R2. We observed that LHS was superior to random sampling in both R2 average and confidence interval. We also found that increasing the training and validation sample sizes increased the test R2 when testing against unseen cases. Study results suggest the applicability of machine reinforcement learning to estimate the well output at different ICV settings, where the neural network model depends fully on the real-time well feedback and production measurements. By using a machine learning approach during the operation of a well with multiple ICV settings, it would be feasible to estimate the lateral-by-lateral output at unseen scenarios. Hence, it becomes possible to maximize the well output by using an optimization algorithm to determine the optimal ICV settings.