Evidence-audit extraction matrix for AI applied to stability and security in inverter-dominated power systems: 44 primary studies, screening ladder and reported search

Description

This dataset is the complete set of extraction, screening and grading records behind a review that audits the evidence base for artificial intelligence applied to stability and security in inverter-dominated power systems. Forty-four primary studies, admitted by two published screening predicates, are graded paper by paper on validation fidelity, operating regime, physical consistency and reproducibility. The workbook holds twenty worksheets: one per manuscript table, a per-paper grading sheet behind the method-family matrix, a screening ladder covering all 74 candidates with the ground on which each was admitted or removed, the reported literature search with a verdict for every one of 1,244 screened records, a calibrated abstract-level comparison of the corpus against a random sample of the eligible literature, and two sheets of primary evidence: the search behind every cell of the positioning table in the five compared reviews, and a blind second reader's per-paper extraction of the sentence in which each of 38 studies states its own validation environment and regime treatment, with every disagreement against the published grades. A codebook states the grading vocabulary, the admission predicates and what the records do and do not establish. Every count reported in the article can be recomputed from these records, and every screening decision can be checked against the paper it names.

Steps to reproduce

The records were produced in four stages, and the workbook releases the material behind each one. 1. Search. Three strings were issued to Scopus and OpenAlex over the 2020 to 2026 publication window, and the reference lists of the five compared reviews were chained backwards. Search_Protocol gives every query as issued, the source and the records returned per leg; Search_Flow gives the stage counts. The legs returned 6,938 records, 4,346 unique after deduplication by DOI and then by normalised title, of which 1,274 were topically relevant by the stated rule and 1,244 were screened at title and abstract. 2. Screening. Every candidate was tested against the two admission predicates printed in the article and reproduced in the Codebook. P1 requires the method to learn a model or policy from data. P2 requires inverter-based resources to be physically present or explicitly modelled as converters in scenarios carrying a non-trivial share of the reported evidence, under three tests covering representation, share and object of study. Search_Screen carries a verdict and a ground for all 1,244 records. Screening_Ladder carries the same for the 74 candidates taken to full text, of which P1 removes 9 and P2 removes 20, one could not be obtained in full text and is recorded as unresolved, and 44 are admitted. 3. Extraction and grading. Each of the 44 admitted papers was read as printed and graded on the evidence tier, the five regime axes, the converter-driven stability class and the reproducibility fields, using the vocabulary set out in the Codebook. A grade records what a paper states about itself; a term occurring in a reference list, or in a description of somebody else's work, is not evidence about that paper. Extraction was carried out by a single reader, which is the limitation the article states. 4. Independent check. The 38 of the 44 papers held as a searchable source were copied to a neutral location under opaque, shuffled identifiers and graded again by a reader working blind, given the printed rules alone and never the article, its reference numbers or its grades. Table4_TierEvidence deposits that reader's per-paper extraction with page numbers, their grades, and every disagreement with the grades printed in the article. Agreement is 29 of 38 on the evidence tier and 27 of 38 on the converter-driven class, Cohen's kappa 0.58 and 0.56. The six papers not deposited are named in that sheet. To recompute a reported figure, filter the sheet that carries it: tier shares from Table4_Corpus, regime coverage from Table6_Regime under its S, R, n.r. and n/a legend, reproducibility counts from Table7_Repro, and method-family counts from Table3_Assignment. Every sheet is generated mechanically from the article's own tables, so each cell matches the paper as printed.

Institutions

• University of Exeter

  England, Exeter
• University of Manchester

  England, Manchester

Categories

Licence