Seasonal short-term photovoltaic power prediction based on GSK–BiGRU–XGboost considering correlation of meteorological factors

Abstract The intermittency and randomness of photovoltaic power present different characteristics due to seasonal variations, which greatly affects the reliability of power supply. To boost the prediction accuracy of photovoltaic power, a short-term prediction combination model named GSK–BiGRU–XGboost is proposed. First, the Pearson correlation coefficient is adopted to determine highly-correlated meteorological factors to photovoltaic power to construct the input features. Second, the prediction errors of different single models are compared, and the two, i.e., Bidirectional Gated Recurrent Unit (BiGRU) and Extreme Gradient Boosting (XGboost) that have the smallest errors and lowest correlation are selected to construct the combination model. Third, to achieve an appropriate weight coefficient of the model, an improved gaining sharing knowledge-based algorithm (GSK) based on parameter adaption is designed to optimize it effectively. Fourth, seasonal models and year-round model based on GSK–BiGRU–XGboost are compared to reveal the effect of seasonal characteristics. Finally, the influence of historical meteorological data window with different steps is investigated. To verify the performance of GSK–BiGRU–XGboost, it is compared with different single and combination models under different weather conditions. GSK–BiGRU–XGboost achieves a high prediction accuracy of 97.85%, which is 9.46% and 12.43% higher than its member models, respectively. Besides, GSK can lead to a 1.71% improvement in the accuracy.