Accurate aggregated PV power forecasting is a key requirement for the reliable integration of solar energy into power systems. Traditionally, regional forecasts are derived by upscaling forecasts from a limited set of reference parks, restricting both spatial coverage and forecast quality. Two central challenges arise in this context: the limited availability of measurement data from reference parks, and the quality of existing measurements, which is frequently insufficient. The present study demonstrates that including poor quality reference parks in regional aggregation substantially degrades forecast accuracy, highlighting the relevance of data quality beyond mere data availability.

The conventional treatment of unmeasured sites relies on physical forecasting models that convert NWP variables into power forecasts at virtual reference points. Such models require explicit parameterization of system and environmental properties. This limits their generalization capability, introduces systematic errors particularly under rare meteorological conditions, and does not account for self consumption at installation level.

To address these limitations, this study adapts a framework previously proposed for aggregated wind power forecasting. A multitask multilayer perceptron was trained on historical power measurements of high-quality reference parks, incorporating both NWP inputs and technical master data of each site. The trained model was subsequently applied to virtual reference points via zero shot learning, and the inclusion of the inferred forecasts at the virtual sites showed significant improvements in regional wind power forecasting accuracy.

The present study extends this framework to regional PV power forecasting and, critically, to poor-quality reference parks. Given that poor-quality parks and virtual sites share the same fundamental constraint of missing reliable measurements, the zero-shot learning framework is extended naturally to this setting. A model trained on historical PV power measurements from a subset of reference parks infers power forecasts for poor quality parks from their master data and the local NWP. This treatment eliminates the adverse impact of unreliable data. At the same time, it introduces a more diverse representation of the regional master data and NWP distribution, as the high-quality reference parks alone may not be fully representative of the region.

The proposed zero shot learning approach for poor-quality parks consistently outperforms both the exclusion of these sites from the aggregated PV power forecast and their replacement with physical model estimates. These findings support the conclusion that fixed physical parameterizations are inadequate for capturing site specific generation behavior, and that master data and NWP inputs from a larger and more diverse set of sites lead to a more accurate and representative regional forecast.