Julian Oelhaf
—
ML researcher applying offline reinforcement learning and fault-detection benchmarks to power-grid protection
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
No output metrics are currently available in the bibliometric record, so this axis is unscored. It would measure publication volume, venue types, and yearly productivity across the researcher's full publication history.
Impact—
Impact indicators such as citation counts, h-index, or downstream adoption are absent from the current data, so this axis is unscored. It would measure how widely the two 2026 grid-protection studies are cited and used by others.
Focus0/100
The focus score is 0, consistent with the underlying field distribution: catalogued output splits evenly (0.5/0.5) between two adjacent fields — machine learning for power system protection and applied machine learning for electrical power systems — so no single field dominates. Topically, however, both 2026 papers converge on one problem, grid fault protection, spanning detection benchmarks and reinforcement-learning-based tripping.
Collaboration51/100
Collaboration scores 51 (mid-range), built on 12 distinct coauthors in the current record. Both catalogued 2026 papers are first-author but non-corresponding, consistent with leading analyses within a larger team rather than owning the full authorship chain as corresponding author. Recurring co-author names include Siming Bayer, Alexander Luce, Christian Bergler, Andreas Maier, Georg Kordowich, Paula Andrea Pérez-Toro, Tomás Arias-Vergara, and Johann Jäger.
Activity—
No activity metrics (publication frequency, recent-year cadence) are available in this snapshot, so the axis is unscored. It would measure how steadily the researcher publishes over time and how current the overall output remains.
Reproducibility60/100
Reproducibility averages 60 across the two reviewed papers. The offline reinforcement learning study rates high readiness — Evolve calls it an honest feasibility study on a CIGRE benchmark network, though trained with a single seed. The systematic ML fault-detection evaluation rates low, with flagged risks of cross-validation leakage, a single PowerFactory topology, and no deep-learning baselines.
Research summary
Julian Oelhaf works at the intersection of electrical power systems and applied machine learning. The reviewed catalog holds two first-author arXiv preprints, both dated 2026, and both targeting fault protection in power grids. One is a systematic evaluation of machine-learning methods for fault detection and fault line identification, reporting F1 0.991 on 10 ms measurement windows; Evolve flags it as a single-topology PowerFactory study without deep-learning baselines and with possible cross-validation leakage. The other applies offline reinforcement learning — specifically conservative Q-learning (CQL) — to distribution-grid protection, learning line-selective tripping on a CIGRE benchmark network with a simulated F1 of 0.9738, while showing that non-fault security remains poor; Evolve characterizes it as an honest feasibility study rather than a deployment-ready method. The bibliometric snapshot is correspondingly narrow: a 50/50 field split between power-system protection ML and applied ML for electrical power systems, 12 distinct coauthors, a mid-range collaboration score of 51, and a reproducibility score of 60 averaged across the two reviewed papers. Output, impact, and activity metrics are missing from the current data, so this profile should be read as an early-stage snapshot of an emerging grid-ML researcher.
Collaboration pattern
Current data records 12 distinct coauthors and a mid-range collaboration score of 51. Recurring names include Siming Bayer, Alexander Luce, Christian Bergler, Andreas Maier, Georg Kordowich, Paula Andrea Pérez-Toro, Tomás Arias-Vergara, and Johann Jäger. Both 2026 catalogued papers are first-author but non-corresponding, consistent with leading analyses within a larger machine-learning group rather than running projects as corresponding author. No institutional affiliation is present in the data.
Research keywords
Highlights
Offline RL for distribution-grid protection: line-selective tripping in simulation
Systematic ML benchmark for grid fault detection in 10 ms windows
Contrasting reproducibility practice across the two grid-protection studies
Timeline
- 2026First-author arXiv preprint: a systematic evaluation of machine-learning methods for fault detection and line identification in electrical power grids, reporting F1 0.991 on 10 ms windows.
- 2026First-author arXiv preprint: offline reinforcement learning (conservative Q-learning) for distribution-grid protection, learning line-selective tripping on a CIGRE benchmark network with simulated F1 0.9738.
- 2026Both catalogued outputs land in this single year, both first-author and both arXiv-sourced, marking an early-stage concentration on machine learning for power-grid fault protection.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Profile is name-resolved from public bibliographic data; no institution, country, or ORCID is present, and the name may be shared by others.
- Only two arXiv preprints, both from 2026, are in the reviewed catalog; journal or conference publications, if any, fall outside the current data scope.
- Output, impact, and activity dimensions are unscored because the underlying bibliometric metrics (publication counts, citations, yearly cadence) are not available in this snapshot.
- Radar scores are 0-100 values computed from the available bibliometric data; the focus score of 0 reflects the even 50/50 split across two adjacent fields rather than an absence of work.