Predicting production-rate using wellhead pressure for shale gas well based on Temporal Convolutional Network

Abstract

Accurate production prediction plays a key role in the development and management of reservoirs. Since reservoir parameters are difficult to obtain for the hydraulically fractured wells, it seems very important to use wellhead pressure to predict production rate rather than bottom hole pressure and other parameters. In this paper, a new prediction method based on Temporal Convolutional Network (TCN) is proposed, which can predict production rate only based on wellhead pressure by learning past patterns between the two. The TCN model can adaptively learn past sequences of arbitrary length by adjusting the receptive field, in which the causal convolutions make it more reasonable to capture past dependencies and extract information, and each output of the model is only related to past inputs. With the direct multi-step prediction strategy, the model can learn relationship between past input-output. The grid search method is employed to select the appropriate receptive field and hyperparameters of the model. To validate the proposed method, three different shale gas wells from China are selected for evaluation and verification. The various results all show that TCN model outperforms the existing methods in terms of accuracy and trend, with all MAPEs less than 6%. By the ablation experiments of well 1 and well 2, we found that the learning of different patterns helps the TCN model to predict more accurately.