Retail Demand Forecasting Of Supply Chain Based On Meta-learning And Machine Learning Algorithm

Supply chain is a network chain structure formed by upstream and downstream enterprises involved in providing products or services to end users in the process of production and circulation,and is the basis for the operation of various industries in society.This paper mainly focus on supply chain demand forecasting problem,at present most research is still using one mature method to predict all goods without considering complex business scenarios,the suitability between data and model,as well as vast amounts of data for the pressure of algorithm works efficiently.This paper use meta learning ensemble method and machine learning method,compare many kinds of algorithm of prediction,adapted the business scenario,establish an effective supply chain prediction system,meet the needs of different scenarios predicted that demand for supply chain management to provide accurate and effective awareness,at the optimum cost and ensure the efficiency of production,inventory,sales,transportation and other issues,to meet the needs of society,enterprises and consumers.This paper uses the service data of a supply chain company in the past three years,including the order,commodity,and promotion data of tens of thousands of commodities,and aims to predict the daily granularity of each commodity of store sales in the next 30 days.For data outliers,this paper uses the IQR method and the 3σ method to remove large orders and smooth sales in the order dimension and the sales dimension respectively.Due to the diversity and complexity of supply chain scenarios,adhering to the concept of "divide and conquer",this paper builds a supply chain demand forecasting system based on three dimensions:sales scenario,sales volume,and sales trend.algorithm strategy for processing.First,for conventional forecasting scenarios,this paper innovatively proposes a time series classification model based on STL decomposition and SBA demand classification method,and divides hundreds of thousands of time series data into five categories:seasonality,trend item,volatility,discontinuity,and sparsity.For each type of sequence,combined with the characteristics of the algorithm model,using different base model pools,by constructing effective time series features,a meta-learning classifier is trained,and each base model is given appropriate weights for integrated prediction.The results show that the meta-learning ensemble prediction algorithm(prediction accuracy 1-wMAPE 71.43%)is significantly better than the simple average ensemble prediction method(prediction accuracy 1-wMAPE 34.04%),especially after implementing the time series classification strategy,the prediction accuracy 1-wMAPE is improved by about 10 percentage points.For the promotion prediction scenario,this paper uses the machine learning algorithm LightGBM as a regressor to make predictions,quantifies the discounts of different types of promotions,and builds effective data features to predict the sales during the promotion.The prediction accuracy rate can reach 72.15%.The innovations of this paper are:1)combining the STL decomposition model and SBA demand classification,innovatively proposes a time series classification algorithm,and selects appropriate base model pools for different sequence trend types,and applies them to metalearning ensemble forecasting;2)in the selection of meta-learner,this paper improves XGBoost to LightGBM,and finally improves the prediction accuracy and engineering efficiency to a certain extent;3)when using machine learning algorithms for promotion prediction,this paper uses mixed multi-step forecasting strategy which combining direct multi-step prediction and recursive multi-step forecasting,avoids the loss of promotional features and the problem of error accumulation,and the model results also show that the forecasting effect is good.