A novel wind resource assessment methodology utilizing transfer learning of multi-mast time-series wind speed

Wind resource assessment is critical for wind farm construction, and high-precision spatiotemporal wind speed prediction at any site is a core prerequisite for ensuring the reliability of assessment outcomes. This study proposes a novel transfer learning-based spatiotemporal wind speed prediction method that integrates computational fluid dynamics and multi-meteorological-mast observational data to enable accurate wind speed calculation at arbitrary target sites within wind farms. The method comprises three core steps: generating a spatial wind speed gain ratio dataset via full-field computational fluid dynamic simulations, fusing this dataset with time-series data from a single benchmark meteorological mast to construct a benchmark spatiotemporal wind speed dataset, and establishing a mapping relationship between multi-mast benchmark data and target site observations using transfer learning algorithms. We systematically investigate the effects of key factors—including flow field simulation errors, benchmark mast selection, and the number of integrated masts—on prediction accuracy, validating the method under typical scenarios: different terrain and short-term wind speed observations. Results demonstrate that compared with traditional inverse distance weighting and measure-correlate-predict methods, the proposed method improves wind benchmark assessment accuracy by more than 20% on average, while exhibiting excellent robustness and stability, making it suitable for wind farms with complex terrain or limited measurement durations.