Short-Term Wind Speed Forecasts over the Pearl River Estuary: Numerical Model Evaluation and Deterministic Post-Processing

<p indent="0mm">The Pearl River Estuary (PRE) is one of China’s busiest shipping hubs and fishery production centers, as well as a region with abundant island tourism and wind energy resources, which calls for accurate short-term wind forecasts. First, this study evaluated three operational numerical models, i.e., ECMWF-EC, NCEP-GFS, and CMA-GD, for their ability to predict short-term wind speed over the PRE against in-situ observations during 2018–2021. Overall, ECMWF-EC outperforms other models with an average RMSE of <sc>2.24 m s^–1 </sc> and <italic>R</italic> of 0.57, but the NCEP-GFS performs better in the case of strong winds. Then, various bias correction and multi-model ensemble (MME) methods are used to perform the deterministic post-processing using a local and lead-specific scheme. Two-factor model output statistics (MOS2) is the optimal bias correction method for reducing (increasing) the overall RMSE (<italic>R</italic>) to <sc>1.62 (0.70) m s^–1 </sc>, demonstrating the benefits of considering both initial and lead-specific information. Intercomparison of MME results reveals that Multiple linear regression (MLR) presents superior skills, followed by random forest (RF), but it is slightly inferior to MOS2, particularly for the first few forecasting hours. Furthermore, the incorporation of additional features in MLR reduces the overall RMSE to <sc>1.53 m s^–1 </sc> and increases <italic>R</italic> to 0.74. Similarly, RF presents comparable results, and both outperform MOS2 in terms of correcting their deficiencies at the first few lead hours and limiting the error growth rate. Despite the satisfactory skill of deterministic post-processing techniques, they are unable to achieve a balanced performance between mean and extreme statistics. This highlights the necessity for further development of probabilistic forecasts.