17:15 〜 18:45
[AAS03-P04] Machine learning prediction of the Madden-Julian Oscillation using reservoir computing extends beyond one months
★Invited Papers
キーワード：マッデン・ジュリアン振動、機械学習
This work presents a novel prediction method of the Madden-Julian Oscillation (MJO), a massive tropical weather event in the tropics that has a large socio-economic impact due to its far-reaching influence on the global weather pattern. The phenomenon is recognized to be the most important source of predictability in extended-range weather forecasts and accurately predicting the MJO has been one of the urgent problems for atmospheric modeling. However, predicting the MJO has been both notoriously difficult and computationally expensive with the state-of-the-art physics-based dynamical weather prediction models (dynamical models). The inherent predictability limit of the MJO has been estimated to be about six to seven weeks from dynamical model ensemble studies but performances of neither dynamical models nor machine learning models have reached this limit.
The potential of machine learning models has been explored in recent years to overcome the difficulty of MJO prediction. However, thus far, the MJO forecast skills of machine learning models have been approximately 20 days and are outperformed by dynamical models by approximately a week. In this study, we employ the reservoir computing method, a brain-inspired machine learning technique, to construct a first machine learning MJO prediction model that is both computationally inexpensive and competitive with state-of-the-art dynamical models.
The reservoir computing model of this study was trained to skillfully forecast the time evolution of the real- time multivariate MJO index (RMM), a macroscopic variable that represents the state of the MJO. The prediction skill of our model was extended by the refinement of the training data. A novel filter (realtime bandpass filter) was developed to extract the recurrency of MJO signals from the raw atmospheric data and restrict the degrees of freedom of the training. The efficacy of the reservoir computing was further enhanced by selecting a suitably correlated time-delay coordinate of the RMM for the training. The constructed model demonstrated the skill to predict the RMM sequence for a month from the pre-developmental stages of the MJO, comparable with that of the dynamical models. Furthermore, the best-performing forecasts predicted the RMM time series for more than two months, which is longer than the predictability limit estimates from dynamical models. Our results indicate that some MJO events are inherently predictable for periods longer than have been previously expected and that there is a possibility for significant improvements in dynamical models to extend their lead time in MJO prediction.
The potential of machine learning models has been explored in recent years to overcome the difficulty of MJO prediction. However, thus far, the MJO forecast skills of machine learning models have been approximately 20 days and are outperformed by dynamical models by approximately a week. In this study, we employ the reservoir computing method, a brain-inspired machine learning technique, to construct a first machine learning MJO prediction model that is both computationally inexpensive and competitive with state-of-the-art dynamical models.
The reservoir computing model of this study was trained to skillfully forecast the time evolution of the real- time multivariate MJO index (RMM), a macroscopic variable that represents the state of the MJO. The prediction skill of our model was extended by the refinement of the training data. A novel filter (realtime bandpass filter) was developed to extract the recurrency of MJO signals from the raw atmospheric data and restrict the degrees of freedom of the training. The efficacy of the reservoir computing was further enhanced by selecting a suitably correlated time-delay coordinate of the RMM for the training. The constructed model demonstrated the skill to predict the RMM sequence for a month from the pre-developmental stages of the MJO, comparable with that of the dynamical models. Furthermore, the best-performing forecasts predicted the RMM time series for more than two months, which is longer than the predictability limit estimates from dynamical models. Our results indicate that some MJO events are inherently predictable for periods longer than have been previously expected and that there is a possibility for significant improvements in dynamical models to extend their lead time in MJO prediction.