基于多尺度时序分解与特征融合的电子商务商品需求预测研究

Study on E-Commerce Product Demand Forecasting Based on Multi-Scale Time Series Decomposition and Feature Fusion

摘要: 在电子商务环境下，商品需求预测是支撑库存管理、供应链协同及运营决策的重要基础。然而，电商商品需求数据通常呈现出多时间尺度叠加、非平稳性强以及波动频繁等特点，给传统预测方法带来了较大挑战。针对上述问题，本文提出了一种基于多尺度时序分解与特征融合的电子商务商品需求预测方法。该方法首先对原始需求序列进行多尺度分解，将其拆分为反映不同时间特征的若干子序列；随后，针对各子序列分别构建预测模型，以充分挖掘趋势特征、周期特征及随机扰动特征；最后，通过特征融合机制综合多尺度预测信息，得到最终的需求预测结果。基于真实电子商务数据集开展的实验结果表明，与传统统计方法及典型深度学习模型相比，本文方法在平均绝对误差、均方根误差和平均绝对百分比误差等评价指标上均取得了更优的预测性能，验证了所提出方法在电商商品需求预测任务中的有效性与实用性。

Abstract: Demand forecasting plays a critical role in inventory management, supply chain coordination, and operational decision-making in e-commerce platforms. However, e-commerce demand data are often characterized by multi-scale temporal patterns, strong non-stationarity, and frequent fluctuations, which pose significant challenges to traditional forecasting methods. To address these issues, this paper proposes a novel demand forecasting approach based on multi-scale time series decomposition and feature fusion. Specifically, the original demand series is first decomposed into multiple sub-series at different temporal scales, capturing trend, seasonal, and irregular components. Prediction models are then constructed for each sub-series to fully extract scale-specific temporal features. Finally, a feature fusion mechanism is employed to integrate the multi-scale prediction results and generate the final demand forecast. Experimental results on real-world e-commerce datasets demonstrate that the proposed method consistently outperforms traditional statistical models and representative deep learning approaches in terms of MAE, RMSE, and MAPE. These results confirm the effectiveness and practical value of the proposed approach for e-commerce demand forecasting.