半導體產業晶圓需求預測之研究－類神經網路模型

Abstract

本論文所提出之類神經網路預測模型，旨在解決半導體產業晶圓需求不確定性的問題。由於，目前臺灣半導體產業所需之晶圓材料必須仰賴國外進口，而其採購所需前置時間甚有長達二到三個月之久；另外，晶圓的價格高昂，種類並不多，係局ABC物料管制的A級物料；因此，我國半導體產業確實需要一高精確度的預測方法，以作為訂單作業的依據。 過去，常用的市場調查方法，如果在取樣不夠完整或樣本不具代表性的情況下，會致使預估結果的不準確。尤其是臺灣半導體產業晶圓需求量的預測，除了受歷史實際需求資料的影響，更受市場景氣因素、時間因素等等重要因子的影響；為此，本論文嘗試提出類神經網路預測模型，將定性與定量因子整合並納入預測計算體系中。在台積公司的資料實證中，以類神經網路所建立的預測模式，對單期物料預測甚可達到百分之百的正確度，而總和五期的平均誤差率亦有達2%者，實證了本晶圓需求預測模型的優異性。

The Neural Forecasting Model proposed is designed for the uncertainty nature of the wafer demand issue in the semiconductor industry. Currently, the needed primary material, the wafer, is resorted to be imported overseas. The lead-time of purchasing wafer is about 2 to 3 months. Besides, in the Fab, the expensive wafers, with few varieties, should be control under suitable inventory. So, it is urgent for the semiconductor industry to have a precise, simple forecasting method for ordering. In the past, conducting marketing survey has been the most regular method for demand forecasting . But it isn’t satisfactory when the samples aren’t enough or aren’t representative. The wafer demand forecasting in Taiwan is especially effected by factors, such as historical usage amounts, the international marketing environment, seasonal factors etc.. Combining both quantitive and qualitive factors into a forecasting model, this neural forecasting model forms an integrated forecasting computing system. Finally, in this case study, some of the performance of the neural wafer forecasting model is verified with a nearly 0% error rate in one single forecasting, and a 2% error rate in the total of five forecasting periods.