A hybrid graph attention network based method for interval prediction of shipboard solar irradiation

Abstract

Solar energy ships came into being to reduce carbon emissions of global shipping and fossil fuels consumption. Accurate prediction of photovoltaic power generation can improve the economical operation of solar energy ships and reduce the power fluctuation of ship power systems. This study proposes a novel hybrid method to forecast ultra-short-term solar irradiation in a solar energy ship along the east coast of China. An improved graph attention framework is trained to describe the dynamic spatial-temporal topology among five weather stations and a solar energy ship. Initial interval prediction results are then determined by an improved Divided Period Optimization (DPO) method. According to distribution characteristics of the prediction error of shipboard solar irradiation, a Dynamic Prediction Interval Modification (DPIM) method is proposed to further optimize the prediction accuracy and interval width of ultra-short-term solar irradiation. Comparison experiments show that the dynamic connection matrix effectively improves the prediction accuracy, and the DPIM method reduces the PI width of 28.84%–37.79 % in shipboard solar irradiation. The proposed prediction method can accurately predict the shipboard solar irradiation at the period of significant disturbance, and provide effective technical support for the economic scheduling of solar energy ships.