Our Global Forecasts
October 1, 2022
At RedSky AI we are making rapid progress with our data-driven global weather forecasts. The forecasts are generated using a custom designed neural network which is trained on a single year of global historical weather data. The diagnostics shown below provide preliminary comparisons between our forecasts and ERA5 reanalysis (historical estimate of the observed weather) produced by ECMWF which is used as the ground truth. Additionally, we provide a comparison with the Global Forecast System (GFS) weather forecasts produced by the United States' National Weather Service, which is one of the best global forecasts. We focus on the 850hPa pressure level, equivalent to approximately 1.5km above sea level, a key height for assessing the quality of global numerical weather prediction models.
The animation to the left shows wind speed in meters per second on the 850hPa surface for every 6 hours over a 10-day period commencing on the 1st of March 2019. ERA5 reanalyses are shown in the top panel, our forecasts in the middle panel, and the difference between the two in the bottom panel. The neural network produces a realistic forecast across the 10 days, accurately forecasting the development of a tropical cyclone in the Southern Indian Ocean, and producing a realistic series of low pressure systems in the North Atlantic Ocean.
The animation to the left shows temperature in °C on the 850hPa level for every 6 hours over a 10-day period commencing on the 1st of March 2019. ERA5 reanalyses are shown in the top panel, our forecast in the middle panel, and the difference between the two fields in the bottom panel.The neural network produces a realistic forecast across the 10 days with differences between reality and the forecast growing with time as expected.
The global mean root mean square error for wind speed in metres per second at 850hPa. The results were averaged over 10--day forecasts for each of the 31 days in March 2019. The growth with forecast hour for GFS is shown in blue, while the same metric for the neural network is shown in orange. In both cases, ERA5 is used as the ground truth for the observed wind speed. Lower root mean square errors indicate superior model performance.
The global mean root mean square error for temperature in °C at 850hPa. The results were averaged over 10-day forecasts for each of the 31 days in March 2019. The growth with forecast hour for GFS is shown in blue, while the same metric for the neural network is shown in orange. In both cases, ERA5 is used as the ground truth for the observed temperatures. Lower root mean square errors indicate superior model performance.
The animation to the left shows wind speed in meters per second on the 850hPa surface for every 6 hours over a 10-day period commencing 4th of March 2019, centered on near Madagascar. ERA5 reanalysis are shown in the top panel (ground truth), our forecast in the middle panel, and the GFS forecast initialized at the same time on the bottom panel. While both GFS and our neural network based forecasts accurately predict the track of Tropical Cyclone Haleh, seen on the right of the image, GFS does not forecast the development of the tropical Tropical Cyclone Idai, one of the deadliest tropical cyclones to impact Africa. The neural network and ERA5 show the strengthening of Tropical Depression Idai into a tropical cyclone from the 10th of March over the Mozambique Channel. While the neural network does not accurately forecast the track of the cyclone after it develops, it is impressive that it can accurately forecast intensification eight days into the forecast.
The diagnostics presented above demonstrate that our neural network produces forecasts that are highly competitive with other traditional numerical weather prediction models. While our neural network-based forecast model is still in development, it is important to emphasize that our model runs thousands of times faster than a global model such as GFS which requires a supercomputer to run. Our neural network-based forecast model can run on a standard desktop computer. This computational efficiency will allow us to produce much higher resolution forecasts than standard NWP models, and to produce large ensembles of forecasts, so we can fully quantify uncertainties. Stay tuned for future updates. We will also be releasing the first version of our API soon to allow our clients to access our forecasts.