應用機器學習與時間序列建構總體需求預測模型－以某化工業為例

Constructing the Aggregated Demand Forecasting Model by Using Machine Learning and Time Series-Case Study in Chemical Industry

陳彥翰(Yen-Han Chen)

指導教授 ： 江瑞清

中原大學/電機資訊學院/工業與系統工程研究所/碩士(2020年)

https://doi.org/10.6840/cycu202000271

摘要

需求預測為生產計劃中很重要的一部份，以個案公司為例，該公司產品約有二百多項產品，但每項產品之需求量皆不同，因此該工廠生產模式為先預測下一期某產品族之總體需求，再根據預測之總體需求，進行生產排程，進行物料需求規劃後，最終投入生產，屬於存貨型生產(Make-to-stock)。由於該公司所預測之下一期需求與實際需求有落差，導致主生產排程需要時常更動，也會導致物料需求規劃更動，造成人力上與時間上的浪費，因此本研究使用The Holt-Winters Method、季節性自我迴歸整合移動平均模型(Seasonal Autoregressive Integrated Moving Average Model; SARIMA)與支援向量迴歸(Support Vector Regression; SVR)進行預測，最終建立一支援決策預測系統。透過該支援決策預測系統，供高階主管作為總體生產規劃之參考，經由個案公司之數據驗證，The Holt-Winters Method於驗證集之平均絕對值誤差率(Mean Absolute Percentage Error; MAPE)為26.78%，SARIMA於驗證集之MAPE為58.20%，SVR於驗證集之MAPE為36.16%，該公司說明MAPE小於30%為可接受，因本研究所提出之支援決策預測模型中的The Holt-Winters Method MAPE為26.78%，因此判斷本研究所提出之支援決策預測模型為可行。

關鍵字

需求預測 ； 機器學習 ； 支援向量迴歸 ； 時間序列 ； The Holt-Winters Method ； SARIMA

並列摘要

Demand forecasting is a very important part of the production plan. Taking the case company as an example, the company's products have more than 200 products, but the demand for each product is different. Therefore, the factory's production model is to forecast the future aggregated demand of a product group, and then according to the forecast aggregated demand, production scheduling, material requirement planning, and finally put into production, which belongs to Make-to-stock production. Due to the gap between the company’s forecast and the actual demand, the master production schedule needs to be changed frequently, and the material requirement planning will also be changed, resulting in a waste of manpower and time. Therefore, this study applies The Holt-Winters Method, Seasonal Autoregressive Integrated Moving Average Model (SARIMA) and Support Vector Regression (SVR) for forecasting demand, and finally establish a support decision-making forecasting model. Through this support decision-making forecasting model, for senior executives as a reference for the aggregated production planning. Through the data verification of the case company, the MAPE of The Holt-Winters Method in the verification set is 26.78%, and the MAPE of SARIMA in the verification set is 58.20%, The MAPE of SVR in the verification set is 36.16%. The company claimed that the MAPE is less than 30% is acceptable. The MAPE of The Holt-Winters Method in the support decision prediction model proposed by this research is 26.78%. Thus, the proposed support decision-making forecasting model is feasible.

並列關鍵字

Demand Forecasting ； Machine Learning ； Support Vector Regression ； Time Series ； The Holt-Winters Method ； SARIMA

參考文獻

中文參考文獻

何應欽，2007，“作業管理”，第九版，原作者:William J. Stevenson，滄海書局。

吳冠毅，2018，“分類式支援向量迴歸結合例子群演算法預測YOUBIKE車位與車輛數”，臺北市立大學資訊科學系研究所碩士論文。

李曜先，2019，“加密貨幣漲跌及趨勢預測-機械學習與時間序列之應用”，元智大學資訊管理學系研究所碩士論文。

李育儒，2018，“以Spark架構預測股票指數之演算法分析”，東吳大學資訊管理學系碩士論文。