A multi-source data-driven approach for online photovoltaic power prediction

Accurate photovoltaic (PV) power prediction, as a prerequisite for safe grid operation, is highly dependent on a large amount of PV data. However, the inaccuracies in a single PV data source may adversely affect the performance of subsequent prediction. In order to provide a comprehensive view of the data and improve the prediction accuracy of long-term PV prediction, a multi-source data-driven approach for online PV power prediction is proposed in this study. Firstly, this study introduces two typical data processing algorithms: the sliding window algorithm and the improved similar-day algorithm. The latter quantitatively analyses the similarity between different moments, which effectively guarantees the similarity of each period. Then, the multi-source data-driven power prediction model is constructed, which integrates the gated recurrent unit model with quantile regression. And the probability prediction of PV power generation is realized by kernel density estimation and the time-varying weight allocation mechanism. Finally, in order to ensure the effectiveness of the long-term prediction, an online learning mechanism is introduced into PV power prediction. The both experimental cases achieve high interval coverage and low bandwidth, demonstrating that the proposed approach exhibits significant improvements in both accuracy and sensitivity.