Dual-channel spatiotemporal graph learning for demand forecasting and constrained scheduling of power grid safety equipment
Demand for power grid safety tools and equipment is spatially heterogeneous and temporally volatile, and is further constrained by periodic dielectric recertification, certificate-linked deployability and work-order-driven consumption. This study develops a domain-oriented and integrative decision-support framework that couples demand forecasting with constraint-aware scheduling. The supply-demand system is represented as an attributed spatiotemporal graph whose node attributes encode maintenance work orders, certification validity, inventory ledgers and meteorological conditions. A Dual-Channel Spatiotemporal Graph Neural Network (DST-GNN) combines a graph attention channel with a dilated causal convolution channel through a node-time gate for multi-step joint forecasting. Scheduling is cast as a Markov decision process and solved by proximal policy optimization under inventory, warehouse-capacity, lead-time and certification constraints, with infeasible actions projected onto the feasible set. On 54 months of operational records from 14 municipal nodes, DST-GNN attains a single-step mean absolute error of 3.21 ± 0.09 and a mean absolute percentage error of 10.3 ± 0.23%, improving on nine forecasting baselines. The resulting schedule reaches a 95.6% fulfillment rate within 2.4% of a linear-programming oracle in total cost, lowers allocation cost by 22.6% and emergency dispatches by 73.9% against rule-based practice, and holds service at or above 87.9% across eight stress scenarios including an out-of-distribution setting.
Authors
- Hancong Huangfu
- Zhaoxiong Guan (ORCID: https://orcid.org/0000-0001-8862-7788)
- Caihua Song (ORCID: https://orcid.org/0009-0001-7511-8658)
- Xinyi Cao
Institutions
- Guangdong Power Grid Company (China) (CN)
Publication Details
- Journal
- Electric Power Systems Research
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1016/j.epsr.2026.114520
- Primary Topic
- Smart Grid and Power Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00