Abstract ID: 016
The value of machine learning to improve seasonal forecasting in mid-latitudes: The example of surface air temperature in central Japan
Lead Author: Pascal Oettli
Center for Environmental Remote Sensing (CEReS), Chiba University, Japan
Keywords: Seasonal prediction, Hybrid prediction, Machine learning, Statistical modeling, Information flow
Abstract: Due to the ocean memory effect, sea-surface temperature anomalies are considered as the main source of seasonal predictability for precipitation and surface air temperature anomalies, at different time lag. In this way, conditions in the equatorial Pacific are known to influence the seasonal air temperature anomalies in numerous regions of the world. Thus, predicting sea-surface temperature conditions a few months ahead helps to estimate the sign of the surface air temperature anomalies.
In the central region of Japan around Tokyo (called the Kanto region), considered as the centre of Japan’s politics and economy, a strong link exists between summer and winter temperatures, and the electric power demand. Knowing the sign and the intensity of surface air temperature anomalies in this region is crucial for power demand forecasting a few months ahead, for good planning of the fuel management and logistics.
Over the years, the SINTEX-F2 (APL–VAiG–JAMSTEC) seasonal prediction system proved its ability to accurately predict sea-surface temperature anomalies few months in advance. Nevertheless, skills drastically drops when it comes to predict surface air temperature anomalies in the mid-latitudes, particularly because the teleconnection patterns are not captured well by the dynamical system.
Taking the Kanto region as a case study, we propose a new type of hybrid prediction system of the surface air temperature which combines dynamical and statistical approaches. In this hybrid system, the statistical component is aimed to restore the teleconnections between sea-surface and surface air temperature anomalies, particularly in mid-latitudes. This component consists of a set of nine (9) different machine learning algorithms, including kernel, tree-based and boosting methods. The dynamical component provides the predictors (i.e., sea-surface temperature anomalies) of the surface air temperature anomalies.
Results show that at 2-month lead the hybrid model outperforms both the persistence and the SINTEX-F2 prediction of surface air temperature anomalies in the Kanto region. This is also true when prediction skill is assessed for each calendar month separately. Despite the model’s strong performance, there are also some limitations, such as the limited sample size, making more difficult to calibrate the statistical model and to reliably evaluate its skill.
Co-authors:
Masami Nonaka (Application Laboratory Research Institute for Value-Added-Information Generation (APL VAiG), Japan Agency for Marine-Earth Science and Technology (JAMSTEC))
Ingo Richter (Application Laboratory Research Institute for Value-Added-Information Generation (APL VAiG), Japan Agency for Marine-Earth Science and Technology (JAMSTEC))
Hiroyuki Koshiba (JERA Co., Inc.)
Yosuke Tokiya (JERA Co., Inc.)
Itsumi Hoshino (JERA Co., Inc.)
Swadhin K. Behera (Application Laboratory Research Institute for Value-Added-Information Generation (APL VAiG), Japan Agency for Marine-Earth Science and Technology (JAMSTEC))
