These multiple-day severe weather forecasts are in the spirit of the convective outlooks issued by the Storm Prediction Center (www.spc.noaa.gov/products/outlook; Edwards 2015) based on predictions of synoptic-scale environments favorable for severe weather. The advantage of using a dynamical convective-scale prediction model on medium-range timescales is that explicit prediction of storms, instead of just environments, potentially can give forecasts of convective modes and specific hazards.
Figure 13. FSS for surrogate severe predictions at different lead times for 00Z initializations of C-SHiELD 2019.
3.4 S-SHiELD Subseasonal-to-Seasonal Prediction
We briefly describe the characteristics of the Tier-2 S-SHiELD configuration, using a 25-km grid designed for climate integrations and for subseasonal and seasonal predictions. S-SHiELD is configured similarly to the 13-km SHiELD, although SHiELD’s two-day relaxation timescale of SSTs in the MLO towards the “frozen anomalies” is extended to 15 days in S-SHiELD. Unlike the vast majority of climate models, S-SHiELD is nonhydrostatic and uses a more sophisticated microphysics which is updated much more frequently. While these features do make S-SHiELD more expensive than analogous 25-km hydrostatic climate models (cf. Murakami et al 2016; Roberts et al. 2018), previous experience with HiRAM (Chen and Lin 2012, 2013; Gao et al 2018) has shown that nonhydrostatic dynamics and better microphysical-dynamical coupling yields a better representation of mesoscale convective systems and in particular of tropical cyclones, a major emphasis of our group’s research.
The MJO plays a major role in subseasonal variability but has been a challenge for many models to predict or even simulate reasonably (Kim et al. 2018). To explore the MJO prediction skill of S-SHiELD we performed 92 40-day predictions, one initialized at 00Z every two days from 1 October 2011 to 31 March 2012, covering the active Dynamics of the MJO (DYNAMO; Yoneyama et al. 2013) observation period. The Real-time Multivariate MJO Index (RMM; Wheeler and Hendon 2004) is calculated for the hindcasts following the methodology of Xiang et al. (2015) and Vitart et al (2017). For each hindcast we compute daily-mean anomalies of outgoing longwave radiation (OLR) and zonal wind at 200mb (U200) and 850mb (U850), averaged between 15S-15N. These forecast anomalies are not bias-corrected since we use observed climatology as reference instead of model climatology. We then subtract the averaged anomalies of the previous 120 days from the total anomalies to remove the signals of interannual and longer time-scale variability; observed anomalies are appended to the anomalies in the hindcast. The normalized U200, U850 and OLR anomalies are then projected onto the pre-computed Empirical Orthogonal Functions (EOFs) from Wheeler and Hendon (2004) to obtain the two RMM indices.
We find that S-SHiELD with the MLO (Figure 14) has good skill (correlation > 0.7) out to 19 days and useful skill (correlation > 0.5) out to 28 days. The RMSE likewise shows similar skill (RMSE < \(\sqrt{2}\) out to 27 days). This skill may not be representative of other time periods given that skill is known to be higher during strong events (cf. Xiang et al 2015) and the period of evaluation is relatively short. However, this result does give us confidence that S-SHiELD simulates the MJO well enough for useful S2S prediction. We plan to expand our evaluation of the MJO in forthcoming work.