边缘智能驱动的电商供应链联邦学习需求预测方法研究

Research on Demand Forecasting Method for Edge Intelligence-Driven Federated Learning for E-Commerce Supply Chain

摘要: 针对电商供应链中各仓储节点对本地化需求预测的迫切需求，以及传统集中式预测方法面临的数据隐私泄露和高延迟问题，本文提出了一种边缘智能驱动的联邦学习需求预测框架(EdgeFL-DP)。该框架将时序预测模型部署于各边缘仓储节点，通过联邦聚合策略实现模型协同优化，在不共享原始数据的前提下提升全局预测精度。具体而言，本文首先构建了基于LSTM和Transformer的双分支时序预测模型作为本地学习器，分别捕获短期波动特征和长期依赖关系；其次，设计了加权联邦聚合算法FedAdapt，根据各节点上报的数据质量指标和分布特征动态调整聚合权重；再次，引入差分隐私机制对模型梯度进行噪声扰动，并结合梯度Top-K稀疏化策略降低通信开销。在基于真实电商场景构建的模拟数据集上的实验表明，EdgeFL-DP的RMSE相比集中式单分支Transformer基线降低了17.8%，相比集中式单分支LSTM基线降低了23.6% (其中双分支架构贡献约6.0%，FedAdapt贡献约7.6%)，同时将数据传输量减少了87.4%，通信延迟降低了65.2%。在隐私预算ε = 8.0下，借助多节点聚合的噪声平均效应，差分隐私引入的精度损失仅为3.7%。

Abstract: To address the urgent need for localized demand forecasting at warehouse nodes in e-commerce supply chains and the data privacy leakage and high latency problems faced by traditional centralized forecasting methods, this paper proposes EdgeFL-DP, an edge intelligence-driven federated learning framework for e-commerce supply chain demand forecasting. The framework deploys time-series prediction models on edge warehouse nodes and achieves collaborative model optimization through federated aggregation without sharing raw data, thereby improving global prediction accuracy. Specifically, we design a dual-branch predictor combining LSTM and Transformer architectures to capture short-term fluctuation features and long-term dependency patterns respectively, a weighted federated aggregation algorithm FedAdapt that dynamically adjusts weights based on locally reported data quality metrics and distribution characteristics, a differential privacy mechanism for gradient perturbation, and a Top-K gradient sparsification strategy to reduce communication overhead. Experiments on simulated datasets based on real e-commerce scenarios demonstrate that EdgeFL-DP reduces RMSE by 17.8% compared to the centralized single-branch Transformer baseline and by 23.6% compared to Centralized-LSTM, with the dual-branch architecture contributing approximately 6.0% and FedAdapt contributing approximately 7.6% of the improvement. Data transmission is decreased by 87.4% and communication latency by 65.2%. Under a privacy budget of ε = 8.0, the accuracy loss introduced by differential privacy is only 3.7%, benefiting from the noise averaging effect across multiple participating nodes.