The uncertainty of large-scale distributed photovoltaic (DPV) systems increases the difficulty in controlling the distribution network. Accurate power forecasting of DPV in substation areas can provide more comprehensive information supports for the optimized operation of active distribution networks. To address the high modeling cost of large-scale DPV power forecasting in substation areas

a dynamic transfer learning method for DPV power forecasting based on multiple predictive scenario models is proposed. First

high-quality data from various substation areas are integrated to construct the source domain sample set

and an Informer-based source domain model that considers different weather forecasting scenarios is introduced to effectively explore the shared mapping features between meteorological conditions and power output under different weather scenarios. Then

to enhance model adaptability

an appropriate source domain model for each substation area is dynamically selected based on the meteorological features of the day to be forecasted. Finally

a transfer learning forecasting technique based on fine-tuning key layer parameters of the source domain model is developed. By dynamically mining the individual output characteristics of each substation area from historical similar days

the method improves the targeting of the forecasting results. The results of forecasting DPV power for 20 substation areas in the same experimental environment show that

compared with individual modeling of each substation area

this method can be adopted to not only improve prediction accuracy but also significantly enhance prediction efficiency

demonstrating its significant advantages in solving the problem of large-scale DPV power forecasting in substation areas.