Unlocking wind energy potential in data-scarce regions: a deep learning framework for strategic climate action in Vietnam’s Central Highlands

Purpose As Vietnam accelerates its transition toward a sustainable energy future under ambitious national strategies like the Power Development Plan VIII (PDP8), a critical barrier persists: the lack of high-fidelity resource maps for strategic planning. This is particularly acute in data-scarce regions with high potential, such as Vietnam’s Central Highlands. This study aims to address this strategic gap by developing and validating a robust deep learning framework to produce an enhanced and detailed high-resolution wind energy map for this region, providing a conservative proof-of-concept for regional energy planning. Design/methodology/approach The framework integrates coarse-resolution daily wind speed data from an ensemble of 11 global climate models with long-term observational records (1985–2023) from a sparse network of meteorological stations. To preserve synoptic extremes critical for wind power, the downscaling was conducted at a daily temporal resolution. A suite of advanced machine learning and deep learning models were systematically trained. To address the limitations of spatial interpolation in complex terrain, the downscaled 10 m wind speeds were extrapolated to a 149 m commercial hub-height using a terrain-aware power law, integrating high-resolution surface roughness data. A technical potential map (GWh/km²/year) was then generated by combining the turbine power curve (Enercon E-101) with a GIS-based availability index (AI), factoring in constraints such as elevation, slope, land cover and urban buffers. Findings Deep learning architectures, particularly the convolutional neural network and XGBoost, significantly outperformed traditional models, achieving R-Pearson correlation coefficients exceeding 0.85 on daily predictions. Validation against ERA5 reanalysis data demonstrated that our ML-downscaled outputs substantially reduced severe negative biases present in ERA5 over complex terrain. The framework generated one of the most detailed wind energy maps for the region, identifying specific “hotspots” capable of generating significant yields (>6.0 GWh/km²/year) while strictly masking physically constrained zones. Projections under future shared socioeconomic pathways indicate that this wind resource potential is expected to remain robust. Originality/value This study could provide a powerful, replicable methodological blueprint for de-risking renewable energy investment in data-scarce regions. By embedding high-resolution spatial covariates (DEM, roughness) into the vertical extrapolation rather than relying solely on 2D interpolation, it delivers an early, scientifically rigorous tool to guide national energy policy, optimize land-use planning and accelerate the sustainable expansion of wind energy for other regions in the country and abroad.