Ultra-short-term PV Forecasting Based on Time Series Representation Learning and NWP Data Integration

Currently, the mainstream photovoltaic(PV) power forecasting methods generally adopt a data-driven approach for modeling. The construction of input features, a key component of these methods, has a significant impact on the prediction results. Existing modeling methods typically use meteorological data and power directly as feature inputs, or employ statistical methods to manually construct features, which introduce some subjectivity. This paper first addresses this issue by adopting time series representation learning methods for automatic feature learning, studying the deterministic and cyclical fluctuations in photovoltaic output to improve prediction accuracy. Furthermore, while using measured meteorological data and power data as prediction inputs, this paper enhances the model's predictive accuracy on a 4-hour ahead time scale by integrating Numerical weather prediction(NWP) data as predictive covariates. Lastly, the model uses the transformer architecture, the decoder part of which uses a one-step output, enhancing the training and inference speed. The results validated through comparative experiments demonstrate that the ultra-short-term PV power forecasting adopted in this paper, which is based on time series representation learning and the NWP data integration, is able to enhance the overall forecasting accuracy, especially the forecasting accuracy on a 4-hour ahead time scale.