An echo state network with attention mechanism for production prediction in reservoirs

Production prediction in petroleum industry plays a significant role in designing the strategy of the exploration and development. However, due to the complex and uncertain underground formations, it is difficult to obtain accurate forecasts. Most of the recent neural network studies on predicting the production rate in reservoirs are limited in optimizing the network models' parameters and reducing the significant computational cost during the training process. To address these issues, we develop a novel echo-state network (ESN) with attention mechanism to forecast the well performance based on production data. Firstly, in the training stage of the ESN model, we used the attention mechanism to extract the relevant features from contextual production rate; secondly, a dimensionality reduction procedure of the reservoir is carried out; thirdly, the generalization capability of the ESN model will be increased significantly by using the regularization constraints; finally, the generic algorithm will be employed in the validation process to optimize the hyperparameters of network. Five cases were carried out to validate the prediction performance of the presented approach. The comparisons with existing deep learning methods and decline curve analysis (DCA) models were implemented. The obtained results indicate that our built model is significantly superior to other deep learning and DCA approaches currently available in the literature. The approach presented in this paper is not only developed to forecast the short-term well performance, but also the production rate reconstructed and predicted using the new method can be employed to estimate missing flow history.