Abstract:
In this study, a U-net deep-learning neural network is utilized to calibrate a 10-m wind forecast, with forecast lead times of 1–7 days over Xinjiang based on the Global Ensemble Forecasting System (GEFS) reforecast dataset from 2000 to 2019. The two conventional postprocessing methods, namely the decaying averaging method and quantile mapping, are employed in parallel for comparison. Results show that raw GEFS exhibits an asymmetric error distribution tending toward positive biases in Xinjiang, with most conspicuous forecast biases distributed in high-altitude regions such as the Tianshan and Kunlun Mountains. Compared with two reference calibrations, the U-net model improves the forecast ability of wind speeds over the whole area, especially over areas where raw GEFS shows large biases such as the Tianshan and Kunlun Mountains. In particularly, the U-net model effectively reduces the positive biases. Furthermore, to analyze error sources, forecast errors are decomposed based on the mean squared errors. Results show that the sequence term remains the main error source after forecast calibrations, which obviously increases with increasing lead time. The multiple calibration methods display different capabilities of ameliorating different error terms. The advantage of the U-net model over two reference forecasts primarily lies in its improvement in terms of the sequence term. After U-net calibration, sequence term slowly grows with increasing lead time. Even for a 7-day lead time, the U-net model reduced the sequence error term by approximately 60% compared with raw GEFS forecasts.