Phase inconsistency as a major source of error in NGFS forecast

Abstract

South Asian monsoon exhibits multiscale spatiotemporal variability. Analyzing the nature and behavior of numerical weather forecast error associated with these space-time heterogeneities will eventually help in improving the models. We investigate the spatiotemporal error characteristics of the National Centre for Medium-Range Weather Forecasting (NCMRWF) Global Forecast System (NGFS) model over South Asian land and ocean separately. Although error grows with lead-time, it saturates within 3–5 days of forecast initiation. The saturated error is only about 15–25% higher than that of day-1, indicating that most of the error accumulates within first 24-h of forecast. Increase in error over oceanic regions is due to an increase in the area with high error at all precipitation ranges with large day-to-day variability. However, over land error growth is primarily confined at locations of high mean precipitation. Decomposition of error arising due to intensity and phase variations reveals that about 90% of it arises from the model’s inability to capture phase of precipitation at various timescales. We show that NGFS cannot capture synoptic scale variations ($$<10$$ day) after day-2. Both the high-frequency (10–20 day) and low-frequency (30–60 day) intraseasonal variations are reasonably predicted up to day-3. At diurnal timescale, NGFS forecasts show a peak in precipitation about 3–6 h prior to that observed, both over land and ocean. Surprisingly, this error does not change with lead-time. Lastly, we show that major error characteristics do not depend on the seasonal mean monsoon rainfall.