A Supply Chain Demand Forecasting Method Based on Federated Learning
Supply chain demand forecasting is essential for operational efficiency and cost control.However,forecasting in uncertain market environments remains challenging.Within the supply chain,different nodes are often managed by different companies,and due to concerns about protecting their interests and privacy,companies are hesitant to share demand data.This reluctance leads to distorted information propagation,exacerbating demand fluctuations and causing the bullwhip effect,which in turn results in inventory issues,production inefficiencies,and increased costs.Traditional demand forecasting methods,such as moving averages,autoregressive models,and exponential smoothing models,often struggle to accurately capture the complex patterns of demand changes,presenting significant forecasting risks.Conversely,neural networks can automatically learn and capture complex data relationships but require substantial training data,and individual node demand forecasting models often exhibit significant prediction errors.To address these challenges,this paper proposes a supply chain demand forecasting method based on federated learning,aiming to ensure that companies in the supply chain do not disclose original demand data while improving forecasting accuracy through collaborative learning.The goal of this study is to develop a highly accurate demand forecasting model capable of reducing safety stock levels and mitigating the bullwhip effect.To achieve this goal,we employ Long Short-Term Memory(LSTM)networks,a type of recurrent neural network tailored for sequential time forecasting,particularly adept at handling long-term dependencies.The proposed approach involves implementing a distributed forecasting model based on LSTM,encompassing a federated learning server and multiple participants(individual retailers).Each retailer meticulously organizes and preprocesses its historical sales data,training an LSTM model on local devices.Subsequently,the locally trained model parameters are transmitted to the federated learning server for aggregation.The server then disseminates the aggregated global model updates back to each retailer,initiating subsequent rounds of local model training.These iterative procedures continue until the model converges or achieves the desired accuracy thresholds.In the numerical experiments,we assume that the market demand for the three retailers follows the same Poisson distribution,and generate corresponding data to simulate real market demand conditions.To add complexity and reflect seasonal or cyclical demand variations,we introduce linear and trigonometric functions.To evaluate demand forecasting methods,we partition the retailers'data into training and testing sets in an 8:2 ratio.We compare three forecasting techniques across the retailers:moving average,the LSTM model,and our proposed federated learning method.Performance is assessed using metrics including mean absolute error,mean squared error,and root mean squared error.The experimental results show that federated learning,by aggregating models without transmitting raw data,successfully learns demand data information from various retailers.This approach mitigates the challenge of poor performance of individual retailer prediction models on unseen data and facilitates the development of more accurate prediction models while safeguarding data privacy.The innovation of this study lies in using federated learning to address data sharing issues in supply chain demand forecasting and extensively studying the impact of different demand forecasting methods on the bullwhip effect.These experimental results validate the practicality and potential of federated learning in the field of supply chain demand forecasting while also confirming the significant impact of training data volume and the number of retailers on the training process.This research provides new insights and methods for supply chain management,bullwhip effect control,and supply chain collaboration.
万方数据
维普
