Explainable temporal relational learning for constraint-aware PQC migration in a constructed energy-sector benchmark
Abstract Post-quantum cryptography migration can depend on the readiness of shared trust, software, control, and recovery services. We evaluate GridPQC-Graph, a temporal relational estimator coupled to constrained migration planning, in a constructed energy-sector benchmark. Six public power-system cases supply physical topology; deterministic cyber-dependency fixtures, archived primitive timings, and a disclosed stochastic label generator supply the remaining evidence. The benchmark contains 769 assets, 5,242 directed relations, and 110,736 controlled trials. Fifty-three time-indexed features feed calibrated histogram gradient boosting. Training, calibration, unseen-topology testing, and later-epoch testing use disjoint rows. Across 20 seeds, the model attains average precision 0.8884 and Brier score 0.1178 on the held-out 300-bus case. Against a newly calibrated logistic comparator, the corresponding paired differences are 0.00415 and − 0.00244; the logistic comparator has lower expected calibration error. Relation-group substitution measures output sensitivity and local ranking stability, without identifying causal effects. A mixed-integer planner enforces maintenance, criticality, privileged-asset, and prerequisite constraints. Its ten-wave mean exposure retirement is 0.3210, with a consequence proxy of 15.895; readiness-first scheduling trades lower consequence for less retired exposure. Optimizer-matched comparisons separate scoring effects from solver choice. These findings establish conditional benchmark performance, not operational effectiveness, field-disruption probabilities, or safe deployment. Formal stochastic models offer complementary assurance when their transition assumptions and probability bounds are independently validated.
Authors
- Ankit Gupta (ORCID: https://orcid.org/0009-0008-8857-0724)
Institutions
- University of North Texas (US)
Publication Details
- Journal
- Discover Artificial Intelligence
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1007/s44163-026-02760-5
- Primary Topic
- Information and Cyber Security
- Type
- article
- Field-Weighted Citation Impact
- 0.00