An ensemble transfer learning strategy for production prediction of shale gas wells

In order to overcome the training data insufficient problem of model for shale gas wells production prediction in new block, this study proposes a transfer learning strategy of improving neural network as the base learner based on the idea of ensemble learning, which is used for shale gas production prediction across formations/blocks. The proposed transfer learning model aims to improve the gas well production prediction performance of new blocks with limited gas well data. The base learner based on improved neural network tries to find the domain invariant feature extraction between source and target blocks through domain adaptation. Bagging algorithm, a parallel ensemble learning method, is used to combine multiple base models to improve the predictive performance of ensemble models. Then, the prediction model trained by the combined data of source and target domain can be directly applied to predict the production of shale gas wells in target domain. The validity of the model was verified on four shale gas well data sets. Results demonstrate that regardless of the degree of domain migration, the transfer learning model proposed in this study can extract domain invariant features by ensemble learning method, overcome the problem of domain migration between source domain and target domain data sets, and significantly improve the production prediction performance of shale gas wells. This work can effectively provide guidance for the production prediction of shale gas wells in new production blocks.