基于注意力机制的CNN-BiGRU短期光伏发电功率预测

Short-term PV Output Forecast of BiGRU Based on the Attention Mechanism

摘要:

精确的光伏发电短期预测在微电网智能能源管理系统中起着至关重要的作用；文章提出一种基于注意力机制的CNN-BiGRU短期光伏发电功率预测模型；其核心思想是通过CNN提取光伏数据的空间特征,把CNN提取的这些空间特征送入到BiGRU神经网络中,利用BiGRU模型捕捉光伏时序数据集的双向信息流,学习光伏特征的动态变化规律,引入Attention机制为CNN-BiGRU的隐藏层输出赋予权重,减少因时序过长造成的信息丢失,并且突出强相关特征的影响,减少弱相关特征的影响。在美国俄勒冈州本德市公开数据集上做了验证,并与BP神经网络、GRU、BiGRU、基于Attention机制的BiLSTM以及基于Attention机制的BiGRU进行对比,实验结果表明所提模型在预测精度上更有优越性。

Abstract:

Accurate short-term forecast of photovoltaic power generation plays an important role in smart energy management system of microgrid. This paper proposes a cnN-BigRU model for short-term pv power generation prediction based on attention mechanism. The core idea is to extract the spatial features of photovoltaic data through CNN and send these spatial features extracted by CNN to BiGRU neural network. BiGRU model is used to capture the bidirectional information flow of photovoltaic temporal data set and learn the dynamic change rules of photovoltaic features. The Attention mechanism is introduced to assign weight to the output of the hidden layer of CNN-BigRU, reduce the information loss caused by the excessively long time sequence, highlight the influence of strong correlation features, and reduce the influence of weak correlation features. The experimental results show that the proposed model has better prediction accuracy than BP neural network, GRU, BiGRU, BiLSTM based on Attention mechanism and BiGRU based on Attention mechanism.