基于抽样卷积交互网络和改进Informer的多元负荷预测方法

Multivariate Load Prediction Method Based on Sampling Convolutional Interaction Network and Improved Informer

针对多元负荷之间耦合关系复杂、波动性和随机性较强的特点,以及已有模型无法兼顾数据的充分挖掘和高效计算的现状,提出了一种基于抽样卷积交互网络和改进Informer的多元负荷预测模型。模型利用抽样卷积交互网络降低多元负荷数据的排列熵,使多元负荷更容易被预测;利用改进稀疏自注意力模块同时提取多元负荷在时间上的时序特征和空间上的耦合特征。通过算例分析可知:所建立的模型既能充分挖掘多元负荷的耦合关系和时序关系,相较于其他模型具有更低误差;能高效训练,相较于传统注意力模型具有更快的训练和预测速度。

Aiming at the characteristics of complex coupling relationship,strong volatility and randomness among multiple loads,and the current situation that existing models can not give consideration to full data mining and efficient calculation,a multiple load forecasting model based on sampling convolution interactive network and improved Informer was proposed.The model used sampling convolution interactive network to reduce the permutation entropy of multivariate load data,making it easier to predict the multivariate load;the improved sparse self-attention module was used to simultaneously extract the temporal and spatial coupling characteristics of multiple loads.The example analysis shows that the established model can fully tap the coupling relationship and time series relationship of multiple loads,and has lower error compared with other models;it can train effectively and has faster training and prediction speed than traditional attention models.