面向智能配送的外卖订单需求预测模型研究——基于随机森林

Research on Takeaway Order Demand Forecasting Model for Intelligent Delivery—Based on Random Forest

摘要: 为提升智能配送系统中外卖订单需求的预测精度以优化运营效率，本研究针对其时序性、非线性及高噪声特性，提出了一种基于随机森林(RF)与直方图梯度提升树(GBDT)的两阶段Stacking集成学习预测模型(RF-GBDT)。本研究基于人工采集的数据进行了系统性特征工程，并创新性地利用GBDT学习RF初步预测的残差以修正系统偏差。实证结果表明，模型有效识别出“时刻”为最关键特征，且RF-GBDT组合模型性能显著优于单一模型，测试集RMSE、MAE和MAPE分别降低了13.9%、19.0%和19.8%，R²提升至0.9016，证实了该框架能为外卖平台提供更精准可靠的需求预测解决方案。

Abstract: To enhance the prediction accuracy of takeaway order demand in intelligent delivery systems and optimize operational efficiency, this study addresses its temporal, nonlinear, and high-noise characteristics by proposing a two-stage Stacking ensemble learning prediction model (RF-GBDT) based on Random Forest (RF) and Histogram-based Gradient Boosting Decision Tree (GBDT). The study systematically performs feature engineering based on manually collected data and innovatively employs GBDT to learn the residuals of preliminary RF predictions to correct systemic biases. Empirical results demonstrate that the model effectively identifies “time of day” as the most critical feature, and the RF-GBDT combined model significantly outperforms individual models, reducing RMSE, MAE, and MAPE on the test set by 13.9%, 19.0%, and 19.8%, respectively, while increasing R² to 0.9016. This validates that the framework can provide more accurate and reliable demand prediction solutions for takeaway platforms.