LSTM and Transformer-based framework for bias correction of ERA5 hourly wind speeds

Reanalysis-derived wind speeds are central to large-scale wind resource assessment (WRA). However, their coarse spatial resolution often introduces significant biases, particularly in complex terrains and coastal areas. A deep learning (DL) framework using LSTMs or Transformers was introduced to correct systematic biases and temporal variability in reanalysis-derived wind speeds by modeling a time-resolved scaling factor, which is then used to adjust ERA5 wind speeds. The proposed framework's spatiotemporal generalization capability was rigorously evaluated using a test set of 170 independent stations distributed throughout Canada in diverse environmental conditions. Results showed that the DL framework outperformed a standard bias correction method based on the Global Wind Atlas. It improved the median wind speed, the temporal variability, and the probability distributions of ERA5 wind speeds in coastal areas and complex terrains. Specifically, in coastal regions, the DL models increased the explained variability of median wind speed by over 70% relative to ERA5. In regions characterized by high surface roughness length, such as forests and urban areas, these models achieved average improvements of more than 10% in MAE and RMSE of the time series. While the DL models performed well in representing the probability distribution of the most typical wind speed values, some challenges remain in improving the distribution of extreme wind speeds. Overall, this framework represents a promising advancement in enhancing the accuracy of reanalysis-derived wind speeds in large-scale WRA. By reducing biases in ERA5 wind data, the DL framework supports more reliable site selection and estimation of long-term energy production. • Introduction of a new deep learning (DL) framework for ERA5 wind speed correction. • DL models learn time-resolved scaling factors to correct ERA5 wind speeds. • Time-invariant scaling factors are ineffective for improving temporal variability. • DL models improved temporal variability and probability distribution. • Challenges remain in accurately representing extreme wind speed distributions.