基于集成学习的生鲜电商水果销量预测研究

The Research on Sales Forecasting of Fresh Food E-Commerce Based on Ensemble Learning

摘要: 生鲜电商在快速发展的同时面临着高损耗率、高物流成本和库存管理复杂等挑战，精准的销量预测对于优化供应链管理至关重要。本论文针对生鲜商品中水果品类的销量预测问题，首先系统性地设计了包含滞后特征、滚动统计特征、时间特征、周期性特征和外部因素特征在内的30维特征体系，充分考虑了促销活动、节假日、气温和极端天气等电商特有的影响因素；在此基础上，采用随机森林、梯度提升决策树和岭回归三种算法，通过简单平均策略进行集成，构建了基于集成学习的预测模型。实验结果表明，集成学习模型的MSE为0.00286，R²为0.9075，MAPE为5.62%，相较于传统时间序列方法和单一机器学习模型在预测精度和稳定性方面均有显著提升。与移动平均方法相比，MSE降低了63.6%，与支持向量回归相比降低了75.1%。研究为生鲜电商平台的库存管理和运营决策提供了有效的技术支持，具有重要的理论价值和实践意义。

Abstract: While fresh food e-commerce is experiencing rapid development, it faces challenges such as high spoilage rates, high logistics costs, and complex inventory management. Accurate sales forecasting is crucial for optimizing supply chain management. This paper addresses the sales prediction problem for fruit categories among fresh products. First, a systematic 30-dimensional feature system is designed, including lag features, rolling statistical features, temporal features, periodicity features, and external factor features, which fully considers e-commerce-specific influencing factors such as promotional activities, holidays, temperature, and extreme weather conditions. Based on this, three algorithms—Random Forest, Gradient Boosting Decision Tree, and Ridge Regression—are employed and integrated using a simple averaging strategy to construct an ensemble learning-based prediction model. Experimental results demonstrate that the ensemble learning model achieves an MSE of 0.00286, R² of 0.9075, and MAPE of 5.62%, showing significant improvements in prediction accuracy and stability compared to traditional time series methods and single machine learning models. Compared with the moving average method, MSE is reduced by 63.6%, and compared with Support Vector Regression, it is reduced by 75.1%. This research provides effective technical support for inventory management and operational decision-making of fresh food e-commerce platforms, offering significant theoretical value and practical implications.