基于Attention-TCN的液化气日订单量预测模型  

Order forecasting model for liquefied petroleum gas daily retail based on Attention-TCN

作　　者：左乘旭 胡文俊[1,2] ZUO Chengxu;HU Wenjun(Shanghai Institute of Computing Technology,Shanghai 200040,China;Shanghai Shenteng Information Technology Company Limited,Shanghai 200040,China)

机构地区：[1]上海市计算技术研究所,上海200040 [2]上海申腾信息技术有限公司,上海200040

出　　处：《计算机应用》2022年第S01期87-93,共7页journal of Computer Applications

摘　　要：液化气(LPG)订单量受气温、节假日等外生变量的影响,为实现对不同类型变量的建模,保证预测的准确性与时效性,提出一种基于注意力-时序卷积网络(Attention-TCN)的液化气订单量预测模型。该模型将分类变量通过嵌入转换为低维向量,与数值变量一同作为时序卷积网络(TCN)模型的输入,并添加了通道注意力以提高准确率。基于真实的液化气订单量数据,对长短期记忆(LSTM)网络、TCN与Attention-TCN模型进行了对比,实验结果表明,相较于LSTM模型,Attention-TCN模型所需的训练时间减少了40%以上,预测结果的均方根误差、平均绝对值误差与平均百分比误差分别降低了5.40%、5.46%与4.74%,在所测试的三个模型中表现最佳。所提出的Attention-TCN模型具有训练耗时短、预测精度高、预测时效性好等特点。The order of Liquefied Petroleum Gas(LPG)is affected by exogenous variables such as temperature and holidays.To realize the modeling of different types of variables and ensure the accuracy and timeliness of forecasting,an order forecasting model for LPG daily retail based on Attention Temporal Convolutional Network(Attention-TCN)was proposed.Categorical variables were converted into low-dimensional vectors through embedding and used as input of Temporal Convolutional Network(TCN)model together with numerical variables,and channel attention was employed to improve accuracy.Based on real LPG order data,comparative experiments were carried out on Long Short-Term Memory(LSTM)network,TCN and Attention-TCN models.The experimental results show that adding channel attention to the TCN model could improve forecasting accuracy with only a small increase in model parameters.Compared with the LSTM model,the training time of the Attention-TCN model was reduced by more than 40%,the Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and Mean Absolute Percentage Error(MAPE)were reduced by 5.40%,5.46%,and 4.74%,respectively,and performed best among the three models.The proposed Attention-TCN model has the characteristics of short training time,high accuracy and good prediction timeliness.

关 键 词：订单量预测 液化气 深度学习 序列建模 时序卷积网络 注意力机制 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]