Deep Learning for Wind Energy Forecasting with Emphasis on Complex Terrain: A Structured Review of Automated Architecture Search, Data Fusion, Uncertainty Quantification, and Explainability

Wind energy forecasting in complex terrain is a critical challenge for reliable power grid integration. This PRISMA 2020-guided structured review examines 134 eligible studies (2018–2026) that apply deep learning (DL) and other neural-network methods to wind speed forecasting, wind power prediction, and wind field downscaling, with particular emphasis on complex terrain. Four methodological pillars are analyzed: automated deep learning (AutoDL/NAS), multi-source data fusion, uncertainty quantification with emphasis on conformal prediction, and explainable AI (XAI). Only 24 studies (18%) explicitly evaluate mountainous, valley, coastal, or urban settings. CNN-based models are the most frequent architecture family, whereas AutoDL/NAS and automated hyperparameter optimization appear in only 11 studies, and no NAS framework has been evaluated in complex terrain. Reported fusion gains are study-specific (from about 13% RMSE reduction to over 50% MAE reduction) and are not directly comparable across datasets, metrics, and baselines. Conformal prediction offers finite-sample coverage guarantees under exchangeability, yet its empirical reliability under terrain-induced non-stationarity has not been evaluated. XAI appears in only five DL studies and is mostly applied post hoc. No study combines three or more pillars. We identify ten challenges and propose ten future directions toward integrated, terrain-aware, uncertainty-calibrated, and interpretable forecasting frameworks.