Improved Surface Wind Speed Forecasts over Beijing-tianjin-hebei Region of China during Spring by Random forests Approach with Sliding-Time-Window and Region Regression

Improved Surface Wind Speed Forecasts over Beijing-tianjin-hebei Region of China during Spring by Random forests Approach with Sliding-Time-Window and Region Regression[1]

Ruixia Zhao¹, Yong¹, Yong Cao¹, Yong Wang²

¹ National Meteorological Centre, Beijing, China

² Nanjing University of Information Science and Technology, Nanjing, China

Strong winds are among the most significant natural hazards, posing great threats to transportation, construction, agriculture, and even the safety of people's lives. Therefore, the accuracy of wind speed forecasting is concerned very much. In our study, machine learning (ML)-based solutions are developed to reduce forecast errors of 10m wind speed produced by the ECMWF’s Integrated Forecasting System (IFS). Two ML approaches, namely decaying averaging method (DAM) and random forest decision trees (RF), are tested at 1985 stations in Beijing-Tianjin-Hebei (BTH) region during the spring of 2021. Considering the importance of computation efficiency in daily operation and the increasing demand for guidance forecast at any specific locations, an so-called sliding-time-window and region regression RF (SR-RF) is designed by pooling the data from sliding-time-windows in recent two years and the whole target region to train the regression models which can be applied at any points within the region. The different models for different lead times of 3h intervals within 72h and 6h intervals within 240h are designed to be daily updated. SR-RF method shows a significant excellent ability to capture the characteristics of IFS errors with 25% to 46% performance improvements in terms of average absolute error (MAE) for all lead times, which is much higher than 7% to 20% improvements of DAM. In particular, the SR-RF method demonstrates its outstanding performance advantages significantly outperforming DAM by dramatically reducing the large forecast errors of IFS over the high-terrain areas in western and northern region, as well as the eastern coastal regions, and overcoming the weakness of excessive strong wind prediction over BTH region in IFS. Furthermore, taking into account of the importance of elevation, latitude, and longitude predictors second only to 10m wind speed in feature importance analysis, and the excellent performance of SR-RF forecast, it is demonstrated that the SR-RF method designed in this study can learn a good knowledge of error distribution characteristics of numerical weather prediction models under different locations and terrain environments, and can learn the downscaling law well while improving the accuracy of prediction.

[1]Supported projects: National Key R&D Program Project (2021YFC3000903), China Meteorological Administration Key Innovation Team Project (CMA2022ZD04)

Author Introduction: Ruixia Zhao, mainly engaged in statistical post-processing of numerical weather prediction model and developing the operational objective global weather forecasting system (CMA-GOWFS).

E-mail: zhaorx@cma.gov.cn / 122323497@qq.com