Improving the Skill of Subseasonal to Seasonal (S2S) Wind Speed Forecasts Over India Using Statistical and Machine Learning Methods

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 4
• Main Authors: Das, Aheli, Reddy, Dondeti Pranay, Baidya Roy, Somnath
• Format: Journal Article
• Language: English
• Published: Wiley 01.12.2024
• Subjects: AI/ML; calibration; statistical methods; subseasonal to seasonal (S2S) scale; wind speeds
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000187

This study demonstrates a framework to improve the skill of raw 10 m wind speed forecasts from numerical models at the subseasonal to seasonal (S2S) time scales. Monthly mean 10 m wind speeds from the ECMWF‐SEAS5 are calibrated using JRA‐55 as reference by employing three statistical methods, bias‐adjustment, quantile‐mapping, and ratio of predictable components (RPC), and four decision tree‐based ML methods, random forest (RF) and light gradient boosting machine (LGBM), RF and LGBM with past observations that is, RF_lags and LGBM_lags respectively, for all 12 months of the year at 1, 2, 3, 4, and 5 months lead time over the homogenous climate zones of India. The quality and skill of raw and calibrated forecasts are evaluated using root mean squared error (RMSE), ratio of standard deviation, and continuous ranked probability skill score (CRPSS). The raw forecasts have large RMSE values, often >1 m/s and mostly do not have any skill. The calibrated forecasts have an RMSE of ∼0.5 m/s, CRPSS ∼0.4, and RMSE ∼0.3 m/s and CRPSS ∼0.7 from statistical and ML‐based methods respectively. The ML‐based methods therefore produce better S2S wind speed forecasts than the statistical methods. This is a timely study with broad impacts especially for the wind energy industry that requires skillful S2S forecasts for financial planning and decision‐making. Plain Language Summary It is very difficult to generate high quality wind speed forecasts in the subseasonal to seasonal (S2S) time‐scale, months ahead of time. Raw numerical weather model S2S wind speed forecasts have high errors and no skill. In this study, we used statistical and machine learning (ML) methods to correct these raw forecasts over India in the S2S scale. The corrected wind speed forecasts have low errors and high skill than the raw forecasts, when compared with the observations. The corrected forecasts using the ML methods are significantly better than the ones corrected using statistical methods. Overall, calibration improves the quality of raw S2S wind speed forecasts and this can have far‐reaching impacts in the wind energy industry operations. Key Points Raw numerical weather model wind speed forecasts in the subseasonal to seasonal time‐scale have high errors and no skill Calibration reduces these errors and improves the forecast skill; machine learning models calibrate better than statistical models In regions with low predictability, inclusion of past observations in the machine learning models enhances the forecast skill