基于改进注意力机制Transformer网络的快消品销量预测方法

Cigarette sales prediction based on improved multi-head self-attention transformer

销量预测能为企业生产计划、仓储运输提供决策支持,使企业能更好地适应市场需求。快消品销售量受众多因素的影响,具有季节性和周期性规律,传统的线性模型难以准确的预测,本文从长时序列预测的视角,运用深度学习理论,提出了一种基于订单时序和订单频率的改进自注意力机制模型(Sequence-Frequency Transformer,SFTransformer)。首先,基于快消品订单数据构建原始数据集,采用time2vec编码处理订单时序信息,并融合订单数据的时序和频率特征在基于时序的订单数据的不同订单频率分别对应不同的注意力头来关注订单数据的订单时序特征和频率特征;使用Transformer模型架构提取特征进行长时序列预测。在真实数据集上进行对比实验,SFTransformer模型在均方误差(MSE)、平均绝对误差(MAE)、均方根误差(RMSE)3项指标上均取得了最佳性能,验证了本文所提方法的有效性。

Sales forecasting can provide decision support for enterprise production planning,warehousing,and transportation,enabling companies to better adapt to market demand.The sales volume of fast-moving consumer goods is influenced by numerous factors(content of factors),and exhibits seasonal and cyclical patterns.Due to the limitations of traditional linear models in accurately predicting sales,this paper proposes an improved self-attention mechanism model based on order sequence and order frequency from the perspective of long time series forecasting,using deep learning theory(Sequence-Frequency Transformer,SFTransformer).Firstly,the original dataset is constructed based on fast-moving consumer goods order data,and time2vec encoding is used to process the order sequence information,integrating the sequence and frequency features of order data.Different order frequencies correspond to different attention heads to focus on the order sequence and order frequency features of the order data.Finally,the Transformer model architecture is used to extract features for long time series forecasting,and comparative experiments are conducted on real datasets.Experimental results demonstrate that the SFTransformer model achieves the best performance in terms of mean squared error(MSE),mean absolute error(MAE),and root mean square error(RMSE),validating the effectiveness of the proposed approach.