Prediction of production indicators of fractured-vuggy reservoirs based on improved Graph Attention Network

Various storage and seepage spaces exist in fractured-vuggy carbonate reservoirs composed of multi-scale dissolution pores and fractures. The frequent regulation of the working system causes nonlinear and unstable production data along with complex water breakthrough characteristics, which lead to difficulty in real-time prediction. Traditional methods based on the water drive curve and autoregressive machine learning ignore the spatial correlation among production wells and local geological characteristics. To address these problems, we propose a model based on improved attention, the spatiotemporal multi-graph convolutional network (STMCN), for production prediction. The unit production wells are abstracted as directed graph network nodes to establish adjacency, connectivity and correlation graphs to extract the spatial and semantic features from different perspectives through graph convolutional networks. To depict the law of fluid movement, we realize the fusion of spatial and semantic information through dynamic routing. Aiming at the autocorrelation characteristics of production sequences, the model uses the self-attention mechanism to capture the dependencies in production sequences. The gating mechanism is designed to achieve dynamic production prediction by adaptively aggregating the spatiotemporal characteristics. This paper evaluates the predictive performance of our model by two real-world datasets of the Tahe Oilfield. The results show that the mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) of our proposed model for the prediction of well A2 were 3.19, 4.66 and 0.07, respectively, which were better than the relatively new baseline model.