Constraint Interpretable Double Parallel Neural Network and Its Applications in the Petroleum Industry

In this article, a constraint interpretable double parallel neural network (CIDPNN) has been proposed to characterize the response relationships between inputs and outputs. The shortcut connecting synapses of the network are utilized to measure the association strength quantitatively, inferring the information flow during the learning process in an intuitive manner. To guarantee the physical significance of the model parameters, the weight matrices are constraint by the square function operator in the non-negative interval. Meanwhile, the sparsity of model parameters has been fully improved by the constraint. Hence, the proposed network can retrieval the critical geophysical parameters sparsely and robustly, through the injection and production signals from the reservoir recovery history. Finally, a synthetic reservoir experiment is elaborated to demonstrate the effectiveness of the proposed method.