晶圓代工廠之需求規劃模式研究

Abstract

需求規劃為供應鏈規劃的最前端，因此需求規劃的品質對整個供應鏈營運的效率有舉足輕重的影響，然而，需求資訊經過供應鏈網路層層傳遞後，往往嚴重扭曲，若再加上人的行為與判斷涉入，常常始得供應鏈的規劃品質嚴重降低，所以，透過有效率且系統化的預測模型來進行需求預測的聚合以及解析，以降低需求的變異並避免長鞭效應，是提昇需求規劃的準確性之重要考量，也是晶圓代工業避免產能不足或過剩的重要因素之一。

在實際情況中，需求量皆為不確定且隨機變動的，因此難以作出準確的預測，而不可信賴的需求資訊也成為需求規劃中之最大憂慮。本研究將透過一個新的需求預測模型，以企業產能規劃策略的觀點進行需求聚合以及解析效應之研究，用來幫助決策者執行由下而上(Bottom-up)、由上而下(Top-down)等等的需求分析，本研究並實際以一晶圓代工產品需求資料為例，透過本研究所提出的解析及聚合模型，用來驗證本論文所提出的模型。

本研究利用平均絕對百分比誤差(MAPE)和泰爾不等係數(TIC)的評估方式來評價本研究的績效，並比較本研究模型應用前後的預測能力分析，發現了本研究模型的預測能力優於傳統的預測方法，就平均絕對百分比誤差而言，本研究模型的誤差值比原本的預測方法低11.72個百分點，就泰爾不等係數而言，本研究模型的預測方向指標U值也比原本的預測方法略優0.045個單位，除了可驗證本研究所提出的模型能有效地改善需求規劃的品質外，更有利於高階主管進行需求規劃、產能規劃及資本資出的決策。惟本研究仍然有其限制，譬如無法將比較資料拉在同一個時間區段上做比較，以及無法克服景氣變化所帶來的影響，因此，本研究建議未來仍可朝建置更符合晶圓代工業的需求預測模型、更多資料維度以及不同時間軸不同資料解析度等方向研究。

With the help of advanced technology, many advanced Supply Chain Planning tools are now available for supply chain management. Demand planning is the very first task for the planning of the entire supply chain network. Sound and realistic forecasting is the key to good planning in any industry. Forecasting is an important aid in effective and efficient demand planning. Unfortunately, the uncertain demand information propagated through the supply chain network often plagues planning quality. Through strategically demand aggregation and desegregation, demand variation can reduced dramatically. Because foundry is a capital extensive business and capacity expansion is the major capital expense. Hence, it becomes imperative to allocate the resources more efficiently in this industry.

Unplanned demand oscillations, including those caused by stock outs, in the supply chain execution process, create distortions that can wreak havoc up and down the supply chain. In this research, a new demand forecasting model has been developed to analyze button-up and top-down demand forecast. Product life cycle disaggregating rules, order/demand forecast disaggregating rules, capacity constraint rules and fair share rules are the key strategies that developed by this research to analyze demand forecast. The practical data in a wafer foundry has been used to proof the performance of this research.

In this research, Mean Absolute Percentage Error and Theil Inequality Coefficient are used as evaluation methods to measure the performance between the original and new forecasting models. Through measuring the deviation of the forecasts in Company A, a wafer foundry, the new forecasting model is proofed to be better than the original one. The results obtained in this research may not represent all the companies in Foundry industry and still has some constraints inside. But the result of this analysis could be an important input for top management to monitor company’s demand planning performance and account manager’s forecasting accuracy.