Abstract: With the rapid development of e-commerce, the volume of express business has increased sharply, which puts forward higher requirements for the operation efficiency of express enterprises. The effective mining of urban network node information is very important for understanding and predicting the express business volume. However, traditional prediction models often ignore the importance of urban network nodes and the complexity of express business data, resulting in insufficient prediction accuracy. In order to effectively mine the node information of the urban network and improve the prediction accuracy of the average daily express business volume, a daily express business volume prediction model based on Convolution-Bidirectional Long Short-Term Memory Neural Network and Autoregressive Integrated Moving Average (Attention-CNN-BiLSTM-ARIMA, AC-BiLSTM-ARIMA) is proposed.
Firstly, the feature derivation strategy is introduced to construct the importance ranking index of city nodes in the urban network from multiple perspectives, and the PCA-entropy weight TOPSIS model is used to analyze the city importance ranking decision, in order to help express enterprises rationally plan the allocation of resources such as manpower and equipment. Secondly, the comprehensive feature extraction method is used to capture the relevant features of the express business volume time series data: the CNN is used to extract the nonlinear features of express business data, the BiLSTM network to extract the bidirectional time series features of express business data, the ARIMA to extract the linear features of express business data, and the Attention mechanism to assign the weights to the corresponding features. Then, AC-BiLSTM-ARIMA daily express business volume prediction model is established for different node cities in the urban network. Finally, according to the real desensitization data of an express enterprise, the simulation experiment is carried out, and the performance of the model is evaluated by calculating the statistical indicators such as Explained Variance Score (EVS), Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and R². It is found that the proposed method can be effectively applied to the actual express business volume prediction and the performance is better than the comparison algorithms.
This paper constructs the city importance index, and combines machine learning and statistics knowledge to provide an improved method for forecasting express business demand. The optimization model has good generalization ability, and the innovation is mainly reflected in the following two points: first, considering factors such as city activity, contribution rate and growth potential, the feature derivation strategy is introduced to construct the city importance index, which provides a new perspective for express enterprise resource planning; the second is to combine the advantages of CNN, BiLSTM and ARIMA to improve the model’s ability to capture linear, nonlinear and time series features, and to improve the accuracy of prediction through weight allocation.
The model can be applied to the daily operation of express enterprises to help them rationally plan human and resource allocation, optimize distribution routes and improve distribution efficiency according to the prediction results.

Key words: urban network node, feature derivation, integrated feature extraction, autoregressive integrated moving average, attention

摘要： 为了有效挖掘城市网络节点信息,提高日均快递业务量预测精度,提出了一种基于注意力机制的卷积—双向长短时记忆神经网络—差分整合移动平均自回归(Attention-CNN-BiLSTM-ARIMA, AC-BiLSTM-ARIMA)的日均快递业务量预测模型。首先基于PCA-熵权TOPSIS模型,引入特征衍生策略进行城市网络节点重要度排序,以期帮助快递企业合理规划快递人力设备等资源配置;然后采用综合特征提取方法,以捕捉快递业务量数据的线性、非线性和双向时序性特征,并采用Attention机制将相应的特征进行权重分配;最后针对城市网络中不同节点城市分别建立AC-BiLSTM-ARIMA日均快递业务量预测模型,并进行仿真对比验证。结果表明,该模型的解释方差(EVS)为0.973,MAE为0.008,RMSE为0.011,R²为0.971,性能均优于对比算法,有助于合理规划配置快递人力和资源,提高快递配送效率。

关键词: 城市网络节点, 特征衍生, 综合特征提取, 差分整合移动平均自回归, 注意力机制