Production Prediction and Liquid Volume Adjustment Based on Spatiotemporal Graph Neural Network

As offshore oil and gas development progressively extends into deepwater and ultra-deepwater regions, accurate prediction of well production has become a critical task. However, in multi-well co-production systems in such reservoirs, conventional time-series models struggle to characterize inter-well dynamic interference, while existing graph neural networks generally lack the capability to model responses to production strategy adjustments. To address this, this study proposes a hybrid model integrating Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks, developing a multi-well production prediction method adaptable to liquid production rate adjustments. Based on numerical simulation data from heterogeneous reservoirs and actual production data, training samples are constructed by combining static geological attributes and dynamic production sequences. During preprocessing, input data are standardized and denoised to enhance model stability and generalization capability. Simultaneously, an adjacency matrix reflecting inter-well dynamic connectivity is established based on Darcy’s law, embedding physical mechanisms into the graph structure learning. The constructed GCN-LSTM model can collaboratively capture spatial dependencies among well groups and temporal evolution characteristics of production dynamics. Experimental results show that the coefficient of determination (R2) on the training, validation, and test sets all approaches 1, with predictions highly consistent with high-fidelity numerical simulation outputs. This verifies the model’s accuracy in dynamic production performance forecasting for offshore oil and gas, providing a reliable core inference module for deepwater digital twin systems and real-time production optimization.