基于GPT的超短期分布式光伏发电功率预测方法

With a lot of distributed photovoltaic (PV) stations connected to the grid, improving the accuracy of power prediction is helpful to safe and economical operation of the power system. However, most existing methods face challenges in dealing with data scarcity and random fluctuations. This paper proposes an ultra-short-term distributed PV power forecasting method based on Generative Pre-trained Transformer (GPT). Firstly, virtual distributed PV power data is generated with different spatial resolutions. Next, the generated data is used to pre-train a Transformer-based model for predicting PV power from 15 minutes to 4 hours. Finally, the model is fine-tuned on real PV stations with a small data volume. The attention mechanism can learn correlations from lots of historical power data by pre-training. Fine-tuning can realize lightweight deployment and accurate prediction on newly built PV stations, effectively dealing with data scarcity. Models with 12.63M and 103M parameters are pre-trained using virtual data, respectively. The experimental results on datasets from 3 locations show that when there is 1 mo real data, the average RMSE decreases by 37.22%, 32.92% and 10.87% compared with the LSTM, linear model, and Transformer-based model, respectively.