基于Luong Attention机制和特征优选策略的超短期负荷预测方法

Ultra-short-term Load Forecasting Method Based on Luong Attention Mechanism and Feature Optimization Strategy

现有超短期负荷预测研究较少考虑到多变量时序数据的特征选择,以及不同输入序列步长对负荷预测的影响程度。针对以上两点,首先通过基于轻量型梯度提升机的嵌入式特征选择算法筛选出影响负荷预测的关键特征,组成优选特征集合。然后,提出一种基于Luong注意力机制的序列到序列门控循环神经网络超短期负荷预测模型,序列到序列门控循环神经网络模型基于编码-解码结构,其输入序列和输出序列都是可变长度的,通过引入Luong注意力机制,突出对负荷预测起到关键影响作用的输入步长信息。算例表明,所提特征选择算法可以有效选择最优特征集合,所提负荷预测模型可以提升模型对输入信息的提取与利用能力,从而提高超短期负荷预测的精度和收敛性能。

The existing research on ultra-short-term load forecasting seldom considers the feature selection of multivariate time series data or the degree of influence of different input sequence lengths on load forecasting.To solve the above problems,the key features that affect load forecasting are selected through the embedded feature selection algorithm based on LightGBM at first,and the optimal feature set is formed.Then,a sequence-to-sequence gated recurrent unit(S2S-GRU)based ultra-short-term load forecasting model is presented,which incorporates the Luong Attention(LA)mechanism.The S2S-GRU model is based on the encoder-decoder structure,and the lengths of both its input and output sequences are variable.Through the introduction of the LA mechanism,the information about the input length which is of significance to load forecasting is focused.Experimental results show that the presented feature selection algorithm can effectively select the optimal feature set,and the corresponding load forecasting model has a better capability to extract and utilize the input information,thus improving the ultra-short-term load forecasting in terms of accuracy and convergence performance.