Accurate forecasts of solar irradiance, photovoltaic (PV) output and wind-power generation are increasingly important for grid balancing, reserve planning, storage operation and electricity-market decisions. This review examines integrated hybrid deep-learning studies published between 1 January 2023 and 9 August 2026 using two linked layers of evidence: an exhaustive eligibility registry and a more detailed nested audit. The database search returned 6,792 records (1,857 from ProQuest and 4,935 from OpenAlex). After deterministic deduplication removed 1,917 records, 4,875 unique records entered title/abstract screening. We excluded 1,620 at that stage and sought 3,255 reports for retrieval and full-text assessment. Of those reports, 990 could not be retrieved (30.4%), 2,265 were assessed, and 1,821 were excluded at full text. This left 444 database-ledger studies. Reconciliation with the separate Other-Methods/Legacy pathway added two distinct studies (R011 and R015); R005 was already represented by OA-1274 and was counted once. The final eligibility registry therefore contains 446 unique studies. Complete study-level extraction and the full Q1–Q10 Quality Gate were available for a 40-study Core40 subset, where 17 studies met the Primary evidence criteria and 23 were classified QA-Hold. The other 406 eligible studies remain in the registry but are not assigned a study-quality class because an equivalent Q1–Q10 appraisal was not available at the analysis freeze. For that reason, detailed findings on architecture, year/domain patterns, missing data and claim strength are reported only for Core40 and are not presented as prevalence estimates for all 446 studies. We also avoid cross-study ranking because the reviewed studies differ substantially in target scale, climate, dataset, forecast horizon, temporal resolution, preprocessing, split design and metric definition.