High resolution hybrid forecast based on the combination of satellite and an All Sky Imager (ASI) network forecasts

A method to combine a highly resolved All sky imager (ASI) network forecast with a satellite based forecast has been developed. The ASI network forecast input is based on the data from the DLR’s Eye2Sky network. This network is installed in North West Germany and includes 29 ASIs, ten Rotating Shadowband Irradiometers (RSIs) and two reference meteorological stations (based on thermal irradiometers) in an extent of 100 km². This network forecast was developed by our colleges from DLR-SF (publication in preparation). It has a forecast horizon of 30 minutes and a step of 1 min with an update of 30 seconds on a domain of 40 km². The second input is based on our operational satellite forecast at DLR-VE and has a horizon of 6 hours with a step and update of 15 minutes. The satellite domain is reduced to the same 40 km² area.

The method consists of three blocks, forecasts homogenization, regression and prediction. In the homogenization block the satellite forecast is interpolated in space and time to the resolutions of the ASI network forecast. We applied linear interpolation for both resolutions as first test case. In the second block, a linear regression is applied to find the optimal weights of the linear combination of the forecast inputs, including a bias term. The regression is based on timeseries extracted from the historical forecasts (features) where the reference is taken from the historical timeseries of ground measurements (samples). Historical data is used in order to indirectly characterize the mean actual local weather conditions on the domain. It is important to note that the regression is performed independently for every lead time. In the third block, we use the optimized weights and biases along with the present (not historical) forecasts to produce the hybrid forecasts. The hybrid forecast resolutions are the same as the ASI based forecast. The output product can be given as maps or timeseries.

For the test case, we are limited from the ASI network side to a dataset of two full months of forecasts (July and August 2020). The highly resolved hybrid forecast was validated against the individual input sources and satellite persistence. We found that this newly developed forecast outperforms the RMSE of persistence and the individual input forecasts for all calculated lead times. It shows an improvement on RMSE of 5.1% to 14.0% with respect to satellite forecasts and 7.6% to 15.1% with respect to the ASI network forecast on lead times going from 5 to 30 minutes. It also shows a lower RMSE under high variability conditions.