Knowledge-augmented energy consumption prediction for discrete manufacturing via causality-informed spatiotemporal graph neural network

Predicting the energy consumption is a crucial approach to optimize energy management for manufacturing enterprises. Due to the multi-internal plants with distributed production processes and dynamic material flows, the energy consumption in discrete manufacturing exhibits complex spatiotemporal patterns, posing challenges for data-driven prediction methods that fail to capture unbiased time dependencies and multi-scale spatiotemporal dynamics. To address these challenges, a Causality-Informed Spatiotemporal Graph Neural Network (CI-STGNN) is proposed. First, a causality discovery module is proposed to mine the knowledge representation of inter-plant energy consumption correlation in discrete manufacturing. Second, a parallel spatiotemporal feature extraction mechanism is proposed to realize the ensembled temporal modeling of the energy consumption with plant-level temporal dependencies and enterprise-level spatial dynamics under discrete manufacturing characteristics. Third, a dual-stage aggregation method with spatiotemporal feature fusion is proposed to learn plant-specific contribution weights for accurate prediction and interpretable energy management. The experiments on real-world industrial datasets demonstrate that CI-STGNN achieves superior performance compared to the state-of-the-art baselines while providing valuable interpretability for energy management decisions.