This paper evaluates the performance of NWMs (neural weather models) in a New Zealand context relative to NWP (numerical weather prediction). To establish baseline performance, we first consider the error for regional mean sea level pressure (MSLP) and surface wind speed across a large number (n=732) of deterministic forecasts from both NWM and NWP relative to ERA5. Comparing the NWMs Pangu-Weather and GraphCast with the IFS dynamical NWP, we find that GraphCast outperforms IFS (albeit slightly), while Pangu-Weather performs slightly better than IFS for wind speed but worse for MSLP. Using ex-TC Gabrielle as a case study, we next evaluate the performance of Pangu-Weather in predicting an extreme weather event in New Zealand. We find that Pangu-Weather outcompetes the NCEP-GFS dynamical NWP model in predicting MSLP and surface wind speed when error is measured relative to ERA5. However when compared against observational station data, NCEP-GFS performs better. Therefore, focusing on errors relative to reanalysis data may overestimate the performance of NWMs that have been trained on reanalysis, even when out-of-sample test periods are used. Our results show promise for future applications of NWMs in New Zealand.
Strong winds affected the Wellington region on 17 September 2023. On the Remutaka summit, both the extreme wind gust peak (101kt or 187km/h) and the frequency of extremely gusty conditions were historically notable. This note presents some characteristics of the event and suggests avenues for future study.
This paper quantitatively compares New Zealand's regional climate projections from the CMIP5 and CMIP6 downscaled datasets produced by NIWA (hereafter NIWA-CMIP5 and NIWA-CMIP6). Given the global challenges posed by climate change, detailed regional projections are valuable for informed planning and decision-making. Our analysis examines key climate variables under two emissions scenarios (moderate and high) and two future periods: mid-century (2031-2050) and end-century (2081-2100). Both CMIP datasets provide broadly similar projections for annual total rainfall, heavy rainfall, wind speed (average and strong), and solar radiation, and agree that the future will be warmer, and generally drier in the northeast and wetter in the southwest. However, the magnitude of these projected climate changes differs. Specifically, NIWA-CMIP6 projects mean air temperatures to be 0.2-0.3 degrees C higher than NIWA-CMIP5 by mid-century and 0.6-0.9 degrees C higher by century's end. For rainfall, NIWA-CMIP6 projects more drying in the east and north of the country than NIWA-CMIP5. For example, NIWA-CMIP6 North Island rainfall climate change signal is projected to be 5.5% drier than NIWA-CMIP5. Relative to NIWA-CMIP5, NIWA-CMIP6 also projects fewer dry days, more hot days in the north, and less drought-prone conditions (based on potential evapotranspiration deficit), although both datasets agree on the general climate change trends of these variables. Seasonal and regional comparisons reveal additional important differences between the datasets; notably, NIWA-CMIP6 projects wetter summers and drier conditions in other seasons under high emissions scenarios at the end of the century. These differences arise from various factors, including bias correction methodology, equilibrium climate sensitivity, regional model differences, and varied emissions scenarios. Notably, differing bias correction methodologies were identified as a key reason for the lower warming projected by NIWA-CMIP5. This analysis offers stakeholders insights for interpreting the updated projections for their regions.
