Shale Gas Production Forecasting with Well Interference Based on Spatial-Temporal Graph Convolutional Network

Abstract

One of the core assumptions of most deep-learning-based data-driven models is that samples are independent. However, this assumption poses a key challenge in production forecasting—performance is influenced by well interference and reservoir connectivity. Most shale gas wells are hydraulically fractured and exist in complex fracture systems, and the neighboring well characteristics should also be considered when constructing data-driven forecast models. Researchers have explored using the graph convolutional network (GCN) to address this issue by incorporating neighboring well characteristics into production forecasting models. However, applying GCN to field-scale studies is problematic, as it requires training on a full batch, leading to gigantic cache allocation. In addition, the transductive nature of GCN poses challenges for direct generalization to unseen nodes. To overcome these limitations, we adopt the graph sampling and aggregation (GraphSAGE) network architecture, which allows training large graphs with batches and generalizing predictions for previously unseen nodes. By utilizing the gated recurrent unit (GRU) network, the proposed spatial-temporal (ST)-GraphSAGE model can capture cross-time relationships between the target and the neighboring wells and generate promising prediction time series for the target wells, even if they are newly drilled wells. The proposed approach is validated and tested using the field data from 2,240 Montney shale gas wells, including formation properties, hydraulic fracture parameters, production history, and operational data. The algorithm aggregates the first-hop information to the target node for each timestep. The encoder-decoder (ED) architecture is used to generate forecasts for the subsequent 3-year production rate by using the 1-year production history of the wells. The trained model enables the evaluation of production predictions for newly developed wells at any location. We evaluate the model’s performance using P10, P50, and P90 of the test data set’s root mean square error (RMSE). Our method preserves the topological characteristics of wells and generalizes the prediction to unseen nodes while significantly reducing training complexity, making it applicable to larger data sets. By incorporating information from adjacent wells and integrating ST data, our ST-GraphSAGE model outperforms the traditional GRU-ED model and shows enhanced interpretability.