Many advanced power system optimization scheduling algorithms are highly dependent on the availability of historical generation data for accuracy. A common challenge faced by new photovoltaic power plants is the scarcity of such historical data. In this study, we introduce a model-aided learning algorithm for generating PV power data. By incorporating the mechanistic model as a priori knowledge into the learning-based approach in order to improve the consistency of the generated PV power generation data with the real-world mechanical model. This improvement is intended to mitigate the problem of de creasing algorithm robustness due to limited data availability Experimental results demonstrate that our proposed method achieves a reduction in the final optimal scheduling cost by compared to the method that only relys on short-term PV generation data for ambiguity set construction.