The increasing integration of renewable energy sources creates a need for reliable forecasts across multiple time horizons. While short-term and day-ahead power forecasting approaches are well established, longer range forecasting of renewable generation remains less explored. This study investigates the potential of seasonal climate forecasts for photovoltaic (PV) generation in Germany with lead times of up to six months.

The approach uses a 30-member seasonal ensemble forecast from the German Climate Forecast System 2.2, including daily global irradiance and temperature. A targeted 10-member ensemble subsample developed by the German Meteorological Service for Germany is evaluated against the original ensemble to assess the benefit of targeted ensemble selection. The daily meteorological forecasts are temporally reconstructed to provide the hourly inputs required by a physical PV model. PV generation is simulated for different module orientations and aggregated according to the spatial distribution and installed capacity of PV systems in Germany. Forecasts are evaluated for Germany as a whole and its four transmission system operator regions using deterministic and probabilistic verification metrics.

The results show pronounced seasonal and regional differences in forecast skill, with the strongest deterministic skill occurring during spring and summer and reduced skill during winter. Targeted ensemble subsampling improves forecast performance in the vast majority of seasons, with particular improvements during spring and summer. However, benefit of subsampling varies substantially across seasons and regions and the full ensemble performs better in some winter periods. Forecast skill also varies substantially between transmission system operator regions. In contrast, differences across lead times are relatively small compared with the pronounced differences between target seasons. Calibration improves the agreement between ensemble spread and forecast error, although the subsampled ensemble remains more under-dispersive than the original ensemble. Although these trends are generally apparent, their statistical significance is limited by the large error bounds.