Subseasonal forecasting of temperature and precipitation over India using Machine Learning approach

Abstract

Subseasonal forecasting (SSF) is the forecasting of the weather parameters two weeks (weather timescale) to two months (seasons timescale) in advance. SSF was considered a ‘predictability desert’ as it is too long for much memory of the atmospheric initial conditions and too short for slowly varying oceanic variability to be felt sufficiently strongly. Moreover, it is a high dimensional problem as it has to consider predictors from atmosphere-land-ocean. Thus, using various parameters as predictors that capture intra-seasonal variability from these three domains, I tried to investigate the weekly forecast of temperature and precipitation at 2- week, 3-week and 4-week forecast horizon over India by a computationally inexpensive ML model-MultiLLR, which prunes out irrelevant predictors and integrates remaining predictors linearly for each target date. After integrating the MultiLLR model with existing physics based dynamical models, the forecast is found to be more skillfull by 41-57% (for temperature) and 178-401% (for precipitation) than the operational dynamical model ERFS currently used by IMD to forecast sub-seasonal climate. It has also been found that, though dynamical models forecast are more skillfull on shorter timescale (week 2), the hybrid approach of MultiLLR comprising of both dynamical model and statistical model shows higher skill of precipitation forecast on extended range time scale (week- 3, week4). However, for temperature prediction, hybrid approach doesn’t give any better prediction than statistical approach