Machine Learning Convolutional Neural Networks ∙Shuffling Well connectivity Reservoir Simulations.

In mature reservoirs, there are hundreds or thousands of producing and injecting wells operating simultaneously, so it is important to understand the impact of injection wells on producers to maintain pressure and control water production. In the case of reservoirs, this same impact aims to prevent possible reactivations or fault fracturing. In this work, we propose a workflow with two strategies, reduced-physics and data-driven modeling, to determine the communication between producing and injecting wells. Both strategies use historical production data, injection rates (inputs), and liquid production rates (outputs). The reduced-physics modeling strategy is based on the Capacitance Resistance Modeling for Producers (CRMP), which calculates the liquid flowrate of the producing well based on the injection rate, productivity index of produce, time constant and the connectivity between injectors and producers. The connectivities are obtained by minimizing the error between the observed and calculated liquid flowrates. The optimization algorithm used is the Sequential Quadratic Programming (SQP). The data-driven modeling strategy is based on Artificial Neural Networks (ANNs), which only use input and output data. The parameters of the neural network, weights, and viéseses, are adjusted during the training process. Three architectures are studied to connect the inputs and outputs: single-layer perceptron, deep learning with multiple layers, and convolutional neural networks. The backpropagation algorithm is used to adjust the weights of the architectures during training. We propose three alternatives for calculating the connectivities after training. The first one is based on the optimal weights. The second one is based on the average error after shuffling the input data, and the last one is based on the gradient importance. Two synthetic models, Two-phases and Brush Canyon Outcrop, are used to validate the proposed workflow. The results show that the connectivities calculated using gradient importance became closer to the connectivities obtained by the capacitance and resistance model when both are compared. In the case of connectivity error between CRMP and other strategies, Backpropagation, Shuffling, and CNN1D, for injector I-1, they are 25%, 16%, and 11%, respectively. Regarding the connectivity of injector, I-3, the error in comparing CRMP versus Backpropagation, Shuffling, and CNN1D is 12%, 12.5%, and 6%, respectively. Similarly, when analyzing the connectivity of injector, I-5, the errors are 11%, 9%, and 8.5%, respectively. In conclusion, we can say that the CNN1D strategy shows a better approximation in calculating connectivities.