PINN-Based Method for Predicting Flow Field Distribution of the Tight Reservoir after Fracturing

Abstract

The physical-informed neural network (PINN) model can greatly improve the ability to fit nonlinear data with the incorporation of prior knowledge, which endows traditional neural networks with interpretability. Considering the seepage law in the tight reservoir after hydraulic fracturing, a model based on PINN and two-dimensional seepage physical equations was proposed, which can effectively predict the flow field distribution of the tight reservoir after fracturing. Firstly, the dataset was obtained based on physical and numerical models of the tight reservoirs developed by volume fracturing. Furthermore, coupling the neural networks and the two-dimensional unsteady seepage equation, a PINN model was developed to predict the flow field distribution of the tight reservoir. Finally, a systematic study was performed concerning the noise corruption levels, training iterations, and training sample size that affect the prediction results of PINN models. Besides, a comparison between PINN and traditional deep neural networks (DNN) was presented. The results show that the DNN model was not only sensitive to noisy data but also more vulnerable to overfitting as the training iterations increase. In addition, the prediction accuracy cannot be guaranteed when the samples are inadequate (<500). In contrast, the PINN model was less affected by noise and training iterations and thus indicates greater stability. Moreover, the PINN model outperforms the DNN model in the case of inadequate samples attributing to prior knowledge. This study confirms that the adopted PINN model can provide algorithmic support for the accurate prediction of flow field distribution of the tight reservoirs.