Which Operating Conditions Are Worth Simulating? Value-Guided Data Acquisition for Security-Constrained Redispatch under Topology Shift

Simulation labels are not equally useful when predictions feed a constrained decision. We study budgeted adaptation of a dispatch surrogate after an unseen transmission-line outage. A five-member multilayer-perceptron ensemble is trained on 220 secure-dispatch solutions from a synthetic six-bus source grid, then adapted using up to 60 labelled cases from a 160-case target pool after line 8 is removed. Random, diversity, uncertainty, and uncertainty-times-predicted-criticality acquisition are compared on the same held-out 80-case target test set across three seeds. At 60 target labels, uncertainty obtains the lowest mean dispatch error (1.387 MW) and repaired-cost regret (21.06 cost units), versus 1.425 MW and 22.42 for the decision-sensitive heuristic. All raw predictions fail an exact 10^-7 MW feasibility test and require a common deterministic repair. A matched-case diagnostic shows a modest topology penalty, and rankings differ across metrics and seeds. The experiment supports decision-oriented evaluation and targeted labelling, but not the claim that the proposed operational weighting improves on uncertainty alone. Code and data: https://github.com/DrUkachi/value-guided-sced