Adaptive masked network for ultra-short-term photovoltaic forecast

Abstract

In recent years, power grid companies have faced increasingly stringent requirements for accurate prediction of photovoltaic (PV) power generation with the rapid development of PV technologies. In ultra-short-term forecasting, PV power generation exhibits strong temporal correlations, leading to high data redundancy. To address this issue, we propose an adaptive masked network (ASMNet) to enhance the accuracy of ultra-short-term PV forecasting. Specifically, this method improves the feature extraction of short-term fluctuations within historical time periods by down-weighting less significant temporal segments during the learning process. It captures the uncertain effects of environmental changes and provides a better understanding of the impacts of ultra-short-term fluctuations. We test our model on three public PV power generation datasets, and it achieves the best performance with a root mean square error of 21.42, 0.2824 and 23.36 for the Belgian, American National Renewable Energy Laboratory, and Desert Knowledge Australia Solar Center datasets, respectively. Additionally, the proposed model demonstrates a 0.01%–0.50% improvement in coefficient of determination compared to baseline models across all datasets, highlighting its superior performance and effectiveness in ultra-short-term PV forecasting.