多种数据补全策略对商超客流量预测影响研究

Research on the Influence of Multiple Data Completion Strategies on the Forecast of Supermarket Passenger Flow

摘要: 数据的准确性对传统线下零售商超的客流量预测有着较为显著的影响。通过对某商超2021年的客流量分析发现，受ERP系统不稳定的影响，传统线下零售商超的客流量曾出现了明显的波动，在2021年3月初到2021年7月末，客流量数据出现了断崖式下降。因此在进行客流量预测时，应当针对数据不稳定性对数据进行一定的特征工程处理。为了探究不同特征工程对客流量预测的影响，我们提出了采用删除异常值、Arima模型填充异常值、Mean-Value填充异常值等多种特征工程处理异常值的方式；使用SArima模型、Holt-Winters模型。在多轮实验中，实验训练集将包含2021年3月初到2021年7月末。并且时间序列模型的参数将以数据为准，选择最优参数进行训练，最终实验生成的模型都将使用2021年8月1日至2021年8月31日的数据做预测验证。结果显示，当不使用任何数据补充的特征工程策略时，Holt-Winters模型较好，而使用了数据补全策略后，SArima模型的预测能力得到了显著提升，并且使用Arima模型进行的数据填充，在一定程度上可以提高SArima模型的预测能力。

Abstract: The accuracy of data has a significant impact on the customer flow prediction of traditional offline retailers. Through the analysis of the passenger flow of a supermarket in 2021, it is found that due to the instability of the ERP system, the passenger flow of traditional offline retailers had a signifi-cant fluctuation, and from the beginning of March 2021 to the end of July 2021, the passenger flow data showed a cliff decline. Therefore, when the passenger flow forecast is carried out, it is neces-sary to carry out some feature engineering processing to the data in view of the instability of the data. In order to explore the influence of different feature engineering on passenger flow prediction, we propose several feature engineering methods to deal with outliers, such as deleting outliers, filling outliers with Arima model, and filling outliers with Mean-Value. Use the SArima model, Holt-Winters model. In multiple rounds of experiments, the experimental training set will cover the period from early March 2021 to the end of July 2021. In addition, the parameters of the time series model will be based on the data, and the optimal parameters will be selected for training. The final models generated by the experiment will use the data from August 1, 2021 to August 31, 2021 for prediction verification. The results show that the Holt-Winters model is better when no data-supplemented feature engineering strategy is used, but the prediction ability of the SArima model is significantly improved after the data completion strategy is used, and the data filling with the Arima model can improve the prediction ability of the SArima model to a certain extent.