Surface temperatures from the ensemble forecasts (ENS) produced by ECMWF’s Integrated Forecasting System (IFS) have often had large winter-time cold biases along the coast in northern Norway. Errors of 10–15°C have not been uncommon in the Lofoten and Tromsø areas. For instance, in Tromsø this winter we saw forecasts below -30°C even though the current minimum record is -18°C.
On 8 March 2016, ECMWF increased the horizontal resolution of its high-resolution (HRES) and ensemble (ENS) forecasts as part of the model upgrade to IFS Cycle 41r2. At the same time an important change was introduced to the radiation scheme. In the new version, the radiative heating and cooling at the surface is improved by approximate updates on the full grid and at every time step (ECMWF Newsletter No. 145, pp. 30–34). In the previous model cycle, 41r1, the radiation scheme was run on a much coarser grid, which contributed to large cold biases at several coastal land points. These model changes were therefore keenly anticipated by MET Norway after reporting back to ECMWF on this issue.
To evaluate the impact of the improved model, we have compared ENS temperature forecasts produced using the two model cycles this winter at 37 coastal weather stations. Probability density functions of the errors for the new and old model cycle show that the old cycle is skewed heavily towards negative values and has a substantial tail of large negative errors. The new cycle is clearly better, with fewer cases of too cold temperature forecasts. Out of 37 stations, 20 have a cold bias exceeding 2°C for the old cycle. This is reduced to 11 for the new cycle. However, some challenges remain: there are still some cases with unrealistic low temperatures. This is particularly true in areas with complex topography near the coast.
How has the impact of the cold bias on forecasts been reduced for users? Medium-range forecasts (3–10 days) are provided as location-specific forecasts on the weather website Yr (www.yr.no). Both consensus and probabilistic forecasts are produced using the ENS 51-member ensemble. Forecast users expect a smooth transition from the short-range forecast, which is based on a 2.5 km convection-permitting model (Arome-MetCoOp in collaboration with the Swedish national meteorological service, SMHI), to the medium-range forecast. To achieve this, ECMWF ensemble re-forecasts are an invaluable tool. The re-forecasts are reruns of the current operational ENS and provide a model climatology. The re-forecasts are particularly important when there are major model upgrades, such as this spring, as they quickly provide a large training dataset suitable for statistical post-processing.
The 2 m temperature ENS forecasts are bias-corrected for Norway using quantile mapping between the re-forecasts and a 2-metre climatology based on a high-resolution model and observations. The probability density function for these reprocessed forecasts shows that this method removes most of the cold bias for the 37 coastal weather stations in northern Norway. The probability density curve is no longer skewed towards negative values. The combination of improved model and statistical post-processing therefore makes us confident that MET Norway can provide its users with good medium-range temperature forecasts, including along the Norwegian coast in winter.