Machine learning prediction of the Madden-Julian oscillation

The socioeconomic impact of weather extremes draws the attention of researchers to the development of novel methodologies to make more accurate weather predictions. The Madden–Julian oscillation (MJO) is the dominant mode of variability in the tropical atmosphere on sub-seasonal time scales, and can promote or enhance extreme events in both, the tropics and the extratropics. Forecasting extreme events on the sub-seasonal time scale (from 10 days to about 3 months) is very challenging due to a poor understanding of the phenomena that can increase predictability on this time scale. Here we show that two artificial neural networks (ANNs), a feed-forward neural network and a recurrent neural network, allow a very competitive MJO prediction. While our average prediction skill is about 26–27 days (which competes with that obtained with most computationally demanding state-of-the-art climate models), for some initial phases and seasons the ANNs have a prediction skill of 60 days or longer. Furthermore, we show that the ANNs have a good ability to predict the MJO phase, but the amplitude is underestimated.

Influence of the Madden–Julian Oscillation on the Arctic Oscillation Prediction in S2S Operational Models

The connections between the Madden–Julian Oscillation (MJO) and the Arctic Oscillation (AO) are examined in both observations and model forecasts. In the observations, the time-lag composites are carried out for AO indices and anomalies of 1,000-hPa geopotential height after an active or inactive initial MJO. The results show that when the AO is in its positive (negative) phase at the initial time, the AO activity is generally enhanced (weakened) after an active MJO. Reforecast data of the 11 operational global circulation models from the Sub-seasonal to Seasonal (S2S) Prediction Project are further used to examine the relationship between MJO activity and AO prediction. When the AO is in its positive phase on the initial day of the S2S prediction, an initial active MJO can generally improve the AO prediction skill in most of the models. This is consistent with results found in the observations that a leading MJO can enhance the AO activity. However, when the AO is in its negative phase, the relationship between the MJO and AO prediction is not consistent among the 11 models. Only a few S2S models provide results that agree with the observations. Furthermore, the S2S prediction skill of the AO is examined in different MJO phases. There is a significantly positive relationship between the MJO-related AO activity and the AO prediction skill. When the AO activity is strong (weak) in an MJO phase, including the inactive MJO, the models tend to have a high (low) AO prediction skill. For example, no matter what phase the initial AO is in, the AO prediction skill is generally high in MJO phase 7, in which the AO activity is generally strong. Thus, the MJO is an important predictability source for the AO forecast in the S2S models.

Contributions of weakly coupled data assimilation-based land initialization to interannual predictability of summer climate over Europe

Land surface is a potential source of climate predictability over the Northern Hemisphere mid-latitudes but has received less attention than sea surface temperature in this regard. This study quantified the degree to which realistic land initialization contributes to interannual climate predictability over Europe based on a coupled climate system model named FGOALS-g2. The potential predictability provided by the initialization, which incorporates the soil moisture and soil temperature of a land surface reanalysis product into the coupled model with a DRP-4DVar-based weakly coupled data assimilation (WCDA) system, was analyzed first. The effective predictability (i.e., prediction skill) of the hindcasts by FGOALS-g2 with realistic and well-balanced initial conditions from the initialization were then evaluated. Results show an enhanced interannual prediction skill for summer surface air temperature and precipitation in the hindcast over Europe, demonstrating the potential benefit from realistic land initialization. This study highlights the significant contributions of land surface to interannual predictability of summer climate over Europe.

Characteristics of Subseasonal Winter Prediction Skill Assessment of GloSea5 for East Asia

In this study, the characteristics of systematic errors in subseasonal prediction for East Asia are investigated using an ensemble hindcast (1991–2010) produced by the Global Seasonal Forecasting System version 5 (GloSea5). GloSea5 is a global prediction system for the subseasonal-to-seasonal time scale, based on a fully coupled atmosphere, land, ocean, and sea ice model. To examine the fidelity of the system with respect to reproducing and forecasting phenomena, this study assesses the systematic biases in the global prediction model focusing on the prediction skill for the East Asian winter monsoon (EAWM), which is a major driver of weather and climate variability in East Asia. To investigate the error characteristics of GloSea5, the hindcast period is analyzed by dividing it into two periods: 1991–2000 and 2001–2010. The main results show that the prediction skill for the EAWM with a lead time of 3 weeks is significantly decreased in the 2000s compared to the 1990s. To investigate the reason for the reduced EAWM prediction performance in the 2000s, the characteristics of the teleconnections relating to the polar and equatorial regions are examined. It is found that the simulated excessive weakening of the East Asian jet relating to the tropics and a failure in representing the Siberian high pressure relating to the Arctic are mainly responsible for the decreased EAWM prediction skill.

Dynamical Seasonal Prediction of Tropical Cyclone Activity Using the FGOALS-f2 Ensemble Prediction System

There is a distinct gap between tropical cyclone (TC) prediction skill and the societal demand for accurate predictions, especially in the western Pacific (WP) and North Atlantic (NA) basins, where densely populated areas are frequently affected by intense TC events. In this study, seasonal prediction skill for TC activity in the WP and NA of the fully coupled FGOALS-f2 V1.0 dynamical prediction system is evaluated. In total, 36 years of monthly hindcasts from 1981 to 2016 were completed with 24 ensemble members. The FGOALS-f2 V1.0 system has been used for real-time predictions since June 2017 with 35 ensemble members, and has been operationally used in the two operational prediction centers of China. Our evaluation indicates that FGOALS-f2 V1.0 can reasonably reproduce the density of TC genesis locations and tracks in the WP and NA. The model shows significant skill in terms of the TC number correlation in the WP (0.60) and the NA (0.61) from 1981 to 2015; however, the model underestimates accumulated cyclone energy. When the number of ensemble members was increased from 2 to 24, the correlation coefficients clearly increased (from 0.21 to 0.60 in the WP, and from 0.18 to 0.61 in the NA). FGOALS-f2 V1.0 also successfully reproduces the genesis potential index pattern and the relationship between El Niño–Southern Oscillation and TC activity, which is one of the dominant contributors to TC seasonal prediction skill. However, the biases in large-scale factors are barriers to the improvement of the seasonal prediction skill, e.g., larger wind shear, higher relative humidity, and weaker potential intensity of TCs. For real-time predictions in the WP, FGOALS-f2 V1.0 demonstrates a skillful prediction for track density in terms of landfalling TCs, and the model successfully forecasts the correct sign of seasonal anomalies of landfalling TCs for various regions in China.

Factors Determining the Subseasonal Prediction Skill of Summer Extreme Rainfall Over Southern China

The occurrence of summer extreme rainfall over southern China (SCER) is closely related with the boreal summer intraseasonal oscillation (BSISO). Whether the operational models can reasonably predict the BSISO evolution and its modulation on SCER probability is crucial for disaster prevention and mitigation. Here, we find that the skill of subseasonal-to-seasonal (S2S) operational models in predicting the first component of BSISO (BSISO1) might play an important role in determining the forecast skill of SCER. The systematic assessment of reforecast data from the S2S database show that the ECMWF model performs a skillful prediction of BSISO1 index up to 24 days, while the skill of CMA model is about 10 days. Accordingly, the SCER occurrence is correctly predicted by ECMWF (CMA) model at a forecast lead time of ~14 (6) days. The diagnostic results of modeled moisture processes further suggest that the anomalous moisture convergence (advection) induced by the BSISO1 activity serves as the primary (secondary) source of subseasonal predictability of SCER. Once the operational model well predicts the moisture convergence anomaly in the specific phases of BSISO1, the higher skill for the probability prediction of SCER is obtained. The present study implies that a further improvement in predicting the BSISO and its related moisture processes is crucial to facilitating the subseasonal prediction skill of SCER probability.

Emerging Skill in Multi-Year Prediction of the Indian Ocean Dipole

The Indian Ocean Dipole is a leading phenomenon of climate variability in the tropics, which affects the global climate. However, the best lead prediction skill for the Indian Ocean Dipole, until recently, has been limited to ~6 months before the occurrence of the event. Here, we show that multi-year prediction has made considerable advancement such that, for the first time, two general circulation models have significant prediction skills for the Indian Ocean Dipole for at least 2 years after initialization. This skill is present despite ENSO having a lead prediction skill of only 1 year. Our analysis of observed/reanalyzed ocean datasets shows that the source of this multi-year predictability lies in sub-surface signals that propagate from the Southern Ocean into the Indian Ocean. Prediction skill for a prominent climate driver like the Indian Ocean Dipole has wide-ranging benefits for climate science and society.

S2S Prediction in GFDL SPEAR: MJO Diversity and Teleconnections

A subseasonal-to-seasonal (S2S) prediction system was recently developed using the GFDL SPEAR global coupled model. Based on 20-year hindcast results (2000-2019), the boreal wintertime (November-April) Madden-Julian Oscillation (MJO) prediction skill is revealed to reach 30 days measured before the anomaly correlation coefficient of the real-time multivariate (RMM) index drops to 0.5. However, when the MJO is partitioned into four distinct propagation patterns, the prediction range extends to 38, 31, and 31 days for the fast-propagating, slow-propagating, and jumping MJO patterns, respectively, but falls to 23 days for the standing MJO. A further improvement of MJO prediction requires attention to the standing MJO given its large gap with its potential predictability (15 days). The slow-propagating MJO detours southward when traversing the maritime continent (MC), and confronts the MC prediction barrier in the model, while the fast-propagating MJO moves across the central MC without this prediction barrier. The MJO diversity is modulated by stratospheric quasi-biennial oscillation (QBO): the standing (slow-propagating) MJO coincides with significant westerly (easterly) phases of QBO, partially explaining the contrasting MJO prediction skill between these two QBO phases.The SPEAR model shows its capability, beyond the propagation, in predicting their initiation for different types of MJO along with discrete precursory convection anomalies. The SPEAR model skillfully predicts the observed distinct teleconnections over the North Pacific and North America related to the standing, jumping, and fast-propagating MJO, but not the slow-propagating MJO. These findings highlight the complexities and challenges of incorporating MJO prediction into the operational prediction of meteorological variables.

NAO Influence on the MJO and its Prediction Skill in the Subseasonal-to-Seasonal Prediction Models

Based on the database of the Subseasonal to Seasonal (S2S) Prediction project of the World Weather Research Programme (WWRP) / World Climate Research Programme (WCRP), the influence of the North Atlantic Oscillation (NAO) on the Madden- Julian Oscillation (MJO) and its forecast skill is investigated. It is found that most models can capture the MJO phase changes following positive and negative NAO events. About 20 days after initialized with a positive (negative) NAO, the forecast MJO appears more frequently in phase 7 (3), which corresponds to reduced (enhanced) convection in the tropical Indian Ocean and enhanced (suppressed) convection in the western Pacific. In most S2S models the MJO prediction skill is dependent on the NAO amplitude and phase in the initial condition. A strong NAO leads to a better MJO forecast skill than a weak NAO. The MJO skill tends to be higher when the forecast starts from a negative NAO than a positive NAO. These results indicates that there is a strong Northern extratropical influence on the MJO and its forecast skill. It is important for numerical models to better represent the NAO influence to improve the simulation and prediction of the MJO.