Long - term wind speed forecasting for the monsoon seasons at station scales over India: Integrating ML and Numerical techniques

Abstract

Long-term wind speed forecasting is still in its early stages, particularly in India. Due to lack of operational forecasts the Indian wind industry is forced to rely on climatological averages, that do not incorporate interannual variability. The overall goal of our study is to evaluate and enhance the capability of the Indian Institute of Tropical Meteorology Coupled Forecast System Version 2.0 (IITM CFSv2) model to forecast the summer monsoon (June-September) 10m wind speeds over India at seasonal scales as a part of the Monsoon Mission III program. The model runs were conducted in hindcast mode for the period 1981-2017. Initially, we conducted a systematic evaluation to assess the quality of the forecasts initialized in February and March for selected stations by comparing them against observations from the Global Summary of the Day (GSOD) dataset. Our findings indicate that the raw forecasts are poor quality with Symmetric Mean Absolute Percentage Error (SMAPE) in the 70% and 90% range. Next, we developed calibration algorithms using ML techniques to improve the quality of the forecasts. Linear Regression, Random Forest, XGBoost, LSTM, Conv-LSTM, GRU were employed as regression models. The outcomes from the best-performing model demonstrate that calibration significantly enhances the quality of the forecasts. After calibration, the mean absolute error (MAE) values typically fall within the range of 0.5 to 0.9 m/s for most stations, though a few stations exhibit values exceeding 1 m/s, in contrast to the raw forecasts where the error range extends from 1.2 to 2 m/s. The SMAPE is reduced to between 30% and 60% after calibration. When compared with 30-year climatology, the calibrated forecasts in 60% of the stations show a positive Root Mean Square Error Skill Score (RMSESS) ranging from 0.01 to 0.3 whereas the scores for the raw forecasts are showing highly negative skill. This study demonstrates that ML based calibration is a promising technique that can significantly improve the quality of numerical model forecasts and perform significantly better than climatology.