Fluctuation Classification and Feature Factor Extraction to Forecast Very Short-Term Photovoltaic Output Powers
Mao Yang, Xiaoxuan Shen, Dawei Huang, Xin Su
Abstract
In recent years, a spurt of photovoltaic power generation has brought certain impact on stability of the power system, which puts forward higher requirements on accuracy of photovoltaic power prediction. Therefore, this paper proposes a hybrid power prediction model based on fluctuation classification and feature factor extraction. First, based on fluctuation characteristics of photovoltaic power, fluctuation classification is applied to forecast power before the day, and weather is divided into complex fluctuation types and simple types. Then, parallel factor algorithm is used to reduce prediction model redundancy, which can reduce high-dimensional numerical weather prediction feature matrix to extract relevant features. Finally, the Long Short-Term Memory (LSTM) deep learning model is used to forecast very short-term photovoltaic power. The proposed hybrid model is compared with other methods, and photovoltaic data from several sites are selected for comparison and validation in this paper. Simulation results show that very short-term prediction method of photovoltaic power proposed in this paper can significantly improve prediction accuracy.