Dynamic Skill Demand Prediction for New Energy Manufacturing Using a Knowledge Graph-Enhanced Spatiotemporal Graph Neural Network
Abstract
Accurate and forward-looking skill demand prediction is critical for workforce planning, curriculum development, and talent cultivation in the rapidly evolving new energy manufacturing sector. Existing approaches often rely on static occupational categories, retrospective skill taxonomies, or purely temporal models and therefore struggle to capture heterogeneous semantic dependencies, temporal evolution, and event-driven demand shocks. This study proposes an integrated framework that combines a point-in-time New Energy Manufacturing Skill Knowledge Graph (NEM-SKG) with a knowledge-graph-enhanced spatiotemporal graph neural network, termed KG-DyPred. The prediction target is redefined as a Composite Skill Demand Index (CSDI), a validated composite proxy of skill-intensity dynamics rather than observed hiring headcount or direct labor-market demand. The NEM-SKG is constructed as a rolling ontology from multi-source heterogeneous data, including job descriptions, patents, course syllabi, industrial policies, and enterprise news. KG-DyPred encodes heterogeneous meta-paths, captures gated spatiotemporal dependencies, and integrates external events through cross-modal attention. A strict temporal-integrity protocol prevents future information leakage. On a leakage-free 6-month-ahead test set (July–December 2025), KG-DyPred achieves a test RMSE of 0.112 ± 0.004, outperforming the strongest matched-input graph baseline TGN (0.126 ± 0.005) with a paired t-test p = 0.0037. A four-origin walk-forward evaluation confirms that the advantage is stable across distant historical windows. Rolling-horizon analysis shows operationally reliable 3–6 month lead indicators; 6–12 month outputs are low-confidence directional hints requiring expert validation. The framework provides a reproducible and interpretable decision-support tool for dynamic skill demand analysis in strategic emerging industries.
