基于多通道输入和PCNN-BiLSTM的光伏发电功率超短期预测

Ultra-short-term Prediction of Photovoltaic Power Generation Based on Multi-channel Input and PCNN-BiLSTM

由于太阳能和天气变量的随机性和不稳定性,光伏发电具有很高的不确定性,准确的光伏发电功率预测对于光伏电站的短期调度和发电计划的运行至关重要。提出一种基于多模式分解、多通道输入、并联卷积神经网络(parallel convolutional neural network,PCNN)和双向长短期记忆网络(bi-directional long short term memory,BiLSTM)的组合预测方法,用于不同天气类型的超短期光伏发电功率预测。首先,由相关性分析算法确定辐照度和温度是对光伏发电贡献最大的2个环境变量,并根据环境因素与光伏功率波动特征的关联性将全年数据划分为4类;其次,使用完全集合经验模态分解、奇异谱分解和变分模态分解对辐照度、温度和光伏发电功率进行分解,以降低原始数据的复杂度和非平稳性,实现不同模式模态分量规律互补;最后,建立基于PCNN和BiLSTM的组合预测模型,使用PCNN提取不同的深度特征,并将PCNN输出的特征融合后输入到BiLSTM中,使用BiLSTM建立历史数据之间的时间特征关系,学习历史数据间的正、反向规律,在时空相关性分析的基础上得到最终光伏发电功率预测结果。实验结果表明,提出的组合预测方法在超短期光伏发电功率预测中具有较高的准确性和稳定性,并优于其他深度学习方法。

Photovoltaic(PV) power generation has high uncertainties due to the randomness and instability nature of solar energy and meteorological parameters.Hence,accurate PV power prediction is essential in the operation of PV power plants for the short-term dispatches and power generation schedules.In this paper,a combined prediction method based on multi-modal decomposition,multi-channel input,parallel convolutional neural network and bi-directional long/short-term memory neural network is proposed for the ultra-short-term PV power generation prediction in different weather types.First,it is determined by the correlation analysis algorithm that radiation and temperature are the two environmental variables that contribute the most to the PV power generation,and the annual data is divided into four types based on the meteorological factors and the fluctuation characteristics of PV power generation;Secondly,the CEEMDAN,SSD and VMD are used to decompose the radiation,temperature and PV power generation under various weather types,in order to reduce the complexity and non-stationary of the original data and realize the complementation between the modal component regular patterns under different modes;Finally,a combined prediction model is built based on the PCNN and the BiLSTM,The PCNN is uesd to extract different depth features,and the features output by PCNN are fused and input into the BiLSTM,which is establish the temporal feature relationship between the historical data,learn the forward and reverse laws between the historical data.The final PV power generation prediction results are obtained based on the analysis of the spatiotemporal correlation.The experimental results show that the proposed combined prediction has high accuracy and stability in the ultra-short-term PV power generation prediction,outperforming the other deep learning methods.