Abstract
Accurate production prediction in the ultra-high water cut stage is crucial for oilfield development. However, uncertainties from operational adjustments, reservoir heterogeneity, and inter-well interference pose significant challenges. Traditional reservoir-engineering methods rely on idealized assumptions and involve high computational costs in numerical simulations. Existing spatiotemporal graph models often separate spatial and temporal modeling, failing to represent reservoir dynamics. Meanwhile, multi-graph frameworks with fixed weights struggle to capture spatial changes during production adjustments. To address these issues, we propose a Spatiotemporal Attention-Enhanced Multi-Graph Convolutional Network (STA-MGCN) for simultaneous multi-well production forecasting. The method first constructs four graphs that encode Euclidean/non-Euclidean features of the well pattern and applies graph convolution to capture spatial interactions. To enrich the temporal context, a hybrid temporal sequence is constructed by integrating historical lag features with real-time controllable variables, and an attention-enhanced Long Short-Term Memory module is designed to dynamically weight each time step in this sequence. Finally, through layer-wise alternating updates of spatial and temporal features and multi-graph fusion, the model achieves true spatiotemporal coupling and adaptively weights the contributions of each graph. The model is evaluated on heterogeneous reservoir data from a composite five-spot pattern. Experiments show that STA-MGCN outperforms baselines, achieves high prediction accuracy and physical consistency in liquid lifting and well shut-in, and captures reservoir heterogeneity and inter-well interference effectively. Its inference speed is three orders of magnitude faster than numerical simulations, enabling efficient real-time decision-making and complex scenario forecasting.