如何基于过去的销售额数据与规律预测未来的销售额，对于商家精细化管理仓储成本、获取更多的利润具有重要意义。本文基于Blending集成学习方法，提出了一种名为NeuralProphet-DeepAR的融合模型，该融合模型通过将NeuralProphet模型与DeepAR模型这两种在建模方法上具有异质性的时间序列基学习器进行融合，以达到更好的销售额预测效果。

本文首先分别介绍了NeuralProphet模型与DeepAR模型的建模原理，并应用两种模型对沃尔玛商品销售额数据集中的三条销售额时间序列进行了实际数据分析。NeuralProphet模型能够较好的捕捉到销售额时间序列中的趋势效应、周期性效应、自回归效应和节假日效应，同时DeepAR模型基于概率预测的方法，也可较好的捕捉到销售额的变化趋势。之后，本文搭建了所提出的基于Blending集成的融合模型，对两种基模型的点预测进行了集成，并进一步基于集成的点预测的残差构造区间预测。实际分析结果表明，融合模型相较于单独使用NeuralProphet模型和DeepAR模型而言，在销售额时间序列上的点预测和区间预测表现更佳。

How to predict future sales amount based on past sales amount data is of great significance for merchants to refine the management of warehousing costs and gain more profits. In this paper, a blending model named NeuralProphet-DeepAR is proposed based on the Blending ensemble learning method, which blends two different time-series based learners named NeuralProphet and DeepAR, which are heterogeneous in modelling methods, to achieve better sales amount prediction.

This paper firstly introduces the modelling methods of NeuralProphet and DeepAR respectively and applies the two models to analyse three sales amount time series of Walmart. The NeuralProphet model captures the trend effect, the seasonal effect, the autoregressive effect and the holiday effect in the time series, meanwhile the DeepAR model, based on probabilistic prediction, can also capture the trend of sales amount series. After that, this paper builds the proposed fusion model based on Blending algorithm, blends the point predictions of the two base models, and further constructs interval predictions based on the residuals of the blending point predictions. The actual analysis results show that the blending model achieve better performance on making point predictions and interval predictions for the sales amount time series compared with the NeuralProphet and DeepAR alone.