快递业在现代物流体系中扮演重要角色，通过连接产业链上下游各个环节，在服务生产、促进消费、保障民生等方面发挥着巨大作用。衡量快递业发展规模的一个重要指标是快递业务量，实现对快递业务量的准确预测，对推动快递行业高质量发展具有重要意义。快递业务量的预测可以为快递基础设施规划、相关扶持政策提供决策参考，另外，快递服务企业能够根据业务量的预测来调整经营决策，如运输车队的调度、是否加大运力投入等，以提升企业的运营管理效率。 当前针对快递业务量预测的研究，包括不同时间尺度下的业务量数据，如日度、月度以及年度的快递业务量预测，不同的样本数据特征所采用的预测方法也不尽相同。本文基于全国快递市场月度业务量数据集，分别建立求和自回归移动平均（ARIMA）模型、指数平滑（ETS）模型以及多层感知机（MLP）模型对快递业务量进行多步预测，并对各预测模型的性能进行综合评价分析。ARIMA模型中，依据样本数据所体现的趋势及季节性特征，确定了合适的差分阶数，并根据信息准则筛选得到最优的预测模型。针对ETS预测模型，比较了加法和乘法两种不同的季节项加入方式对预测效果的影响，同时还考察了阻尼系数的引入在快递业务量预测任务中的表现。此外，利用相空间重构的原理建立监督学习模型来完成对快递业务量的预测，采用多层感知机作为回归器，并通过网格搜索的方法确定了滑动窗口长度的取值。针对多层感知机模型对业务量季节性特征拟合欠佳的问题，利用基于局部加权回归的STL分解法，将业务量序列分解为多个子序列，然后对子序列进行预测后自下而上汇总为最终预测结果。在测试集上的预测性能评价结果显示，改进后的STL-MLP能够给出最优的预测结果，次优模型为ETS模型。最后，在S快递公司的业务量数据集上对三种预测模型进行了进一步的综合比较，结果支持了STL-MLP预测模型性能的优越性。
The express industry plays an important role in the modern logistics system. By connecting the upstream and downstream of the supply chain, it benefits the manufacturing industry, promotes consumption, and further improves people's livelihood. One of the reliable indicators to measure the scale of the express industry is the express business volume. Accurately forecasting the express business volume is of great significance to achieve the High-Quality Development of the express industry. On the one hand, the forecast of express business volume can improve decision-making for facilities planning as well as supportive policies developing. On the other hand, express enterprises can dynamically make their operational decisions based on the forecast of express business volume, such as the truck scheduling, and capacity planning, to enhance operating efficiency. At present, the research on the forecasting of express business volume includes forecasting at the daily, monthly and yearly frequency, and the methods also varies by datasets with different features. Based on the monthly express business volume of China’s market, this paper establishes an autoregressive integrated moving average (ARIMA) model, an exponential smoothing (ETS) model and a multi-layer perceptron (MLP) model to produce multi-step point forecast. The performance of each model is comprehensively evaluated and analyzed. In the ARIMA model, according to the trend and seasonal components of the sample data, the appropriate orders of differencing are determined, and then use information criterion to select the optimal model. For the ETS model, the effects of the additive and multiplicative form of seasonal component were compared, and the performance of the damped trend method was also investigated. In addition, a supervised learning model is established by using the principle of state space reconstruction theory. MLP is used as the regressor, and the length of the sliding window is chosen by the grid search method. Aiming at the problem that the MLP model lacks the ability to fit the seasonal patterns, the STL method based on Loess is used to decompose the series into multiple sub-series, and then the forecasts of sub-series are produced and summed to obtain the final output. The performance evaluation results on the test set show that the STL-MLP model can provide the best forecasts, and the sub-optimal model is the ETS model. Finally, this paper conduct a further comprehensive comparison of the three models on the express business volume dataset of S company, and the results show that the STL-MLP model has its advantage in the task of express business volume forecasting.
