Currently，various statistical and machine models have been widely applied to photovoltaic （PV） power forecasting，but low forecasting accuracy is common with scarce PV historical data.Therefore，a time generative pre-trained transformer （TimeGPT ） is introduced into the short-term forecasting of PV power.Firstly，a time series large model is constructed based on a large-scale and diverse time series dataset of 100 billion data points （such as finance，traffic，banking，network traffic，weather， energy，and healthcare ）.Then，TimeGPT is fine-tuned using a small amount of PV power historical data to adapt to the data distribution and characteristics related to PV power forecasting.TimeGPT is simulated in the PV dataset with user privacy and compared with existing statistical and machine models.By taking case 1 as an example，the mean absolute error （MAE） of TimeGPT is reduced compared with the comparison models when the forecasting step is 1 h.Finally，the conditions for TimeGPT application and the direction of improvement are summarized for its application in new electric power systems.