以混合式類神經網路改善需求規劃績效於晶圓代工業之個案研究

Abstract

供應鏈之規劃是以需求規劃為起點。當需求的規劃與預測能夠越精準則供應鏈管理的運作就能更順暢，供應鏈的績效也能越高。然而，以半導體產業而言長期以來，由於受到景氣循環以及客戶需求之不確定的影響下，產能往往無法被有效利用，相對的也使得投資效益也日益低落。故如何有效預測客戶需求進而提升產能利用率和投資效益是本研究主要探討的問題。 本研究提出以混合式類神經網路為基礎之同步供應鏈架構的需求預測模型，引入及時的市場資訊以及客戶資訊和歷史訂單與業務預測資訊，經過混合式類神經網路前端之SOFM(自組織特徵映射類神經網路)，可有效的將資訊依其時空背景各自群聚分類，此外本研究更進一步的將分群後的知識帶入後端MLP(多層感知機類神經網路)以進行預測。本研究透過一晶圓代工廠之實際資料實證，發現混合式類神經網路與較純MLP類神經網路(資料的輸入非知識的輸入)預測誤差較小，並且同時均優於與原業務預測對照組和複回歸模型組。 與原本的業務預測對照組比較，混合式類神經網路的誤差標準差較之可減少約26%，換句話說可降低約26%的風險。換言之即平均每月每客戶約可減少211片(約當八吋晶圓)的閒置產能；並且減少平均803片(約當八吋晶圓)的訂單損失，換算成美金則分別為US$222438與US$843897。此外本研究中也發現透過同步供應鏈管理之架構下進行產能的同步規劃，可以提升產能利用率和投資效益。

Demand planning is a starting point of the entire supply chain planning activities. Accurate and realistic forecasting will bestead the performance of supply chain management. Unfortunately, the market dynamic and uncertain demand information will highly affect the supply chain planning quality. Besides wafer foundry is such a capital extensive business and capacity expansion is the major capital expense. The more accurate demand planning will be, the more feasible capacity planning will be carried out. In this paper the Sensitivity Demand Planning Process Model of Synchronous Supply Chain Management Architecture based on hybrid artificial neural networks is proposed. This model imports on time market information、customer inventory and historical order information, which are classified by front-end layer SOFM (Self-Organizing Feature Map). Then it puts the classified knowledge into back-end layer MLP (Multilayer Perceptron) to predict the real demand. After experimented with a wafer foundry real data and comparing output results of hybrid artificial neural networks and original forecast, the mean absolute deviation of hybrid artificial neural networks less than the one of original forecast 、 regression and pure MLP model is found. Through this model, the demand forecast will be more accurate and the capacity utilization will be huge increase synchronously through the SSCMA’s process. Finally, investment effectiveness will be largely refined.