Guarantees for Learning-Based Decisions in Distribution Networks Under Model Mismatch: A Unified View and Benchmark
Year: 2026
Type: preprint
Abstract: This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible. Abstract: Guarantees for learned controllers and operating envelopes in distribution networks hold for a property, on a grid model, over a scope of inputs or time. We make this type explicit, map 103 papers, and find most methods move guarantees along one axis. A transfer lemma reduces model changes to discrepancy bounds. On balanced radial feeders without shunts, linearized models are known to overestimate voltages: upper-limit certificates of rules without voltage feedback transfer to nonlinear power flow, lower limits need a margin. We show that on a three-phase line loaded on one phase, mutual coupling reverses this sign on the other phases for typical impedances. In our benchmark, a neural-network controller verified on the linearized model over its input set violates the nonlinear lower voltage limit on 12.6% of it, as does the imitated optimum (12.7%); mismatch-aware retraining is certified everywhere. With 6 input variables the optimum still violates; for a margin-trained network, 20,000 evaluations of a rigorous nonlinear certificate prove 46.0-48.3% of the inputs safe (78.6-79.0% with guided splitting). On the IEEE 123-node feeder, phase coupling alone falsifies linearized export envelopes in 5 of 16 cases. Per-step 95% voltage margins leave 18%-98% of days unsafe across calibration samples.
Author: Shen, Yuan
Institution: University of British Columbia, Okanagan Campus
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234893737
