<p>Digital manufacturing enterprises require high operational agility due to the intricate and dynamically changing nature of their tasks. The implementation of accurate and timely predictions of task bottlenecks is therefore crucial to enhancing overall efficiency. Due to task complexities and dynamic business environments, bottleneck prediction is a challenging issue. This study introduces a novel approach that constructs a task network from extensive data accumulated within a digital enterprise to identify and depict the complex interrelations among tasks. Based on this method, we develop a Bottleneck Spatio-Temporal Graph Convolutional Network (BTGCN) model based on deep learning methods that considers spatial features of the task network and temporal data of task execution and integrates the strengths of GCN and GRU. We find that GCN effectively learns and represents the complex topology of task networks to capture spatial dependencies, while GRU adapts to the dynamic changes in task data, accurately capturing temporal dependencies. Informed by the theory of constraints, the study applies the proposed BTGCN model to the prediction of task throughput bottlenecks in digital enterprises. Experimental results demonstrate that while the model has certain limitations, it can accurately extract spatio-temporal correlations from system data, offering advantages in bottleneck prediction over other benchmark models.</p>