TFT-LCD 產業階層式先進規劃與排程(APS)

《TFT-LCD 產業階層式先進規劃與排程(APS)》由会员分享，可在线阅读，更多相关《TFT-LCD 產業階層式先進規劃與排程(APS)（42页珍藏版）》请在装配图网上搜索。

1、page 1TFT-LCD Research_GroupCopyright 2010 需求不確定下TFT-LCD產業隨機產能規劃 Stochastic Capacity Planning for TFT-LCD Manufacturing under Demand UncertaintyTzu-Li Chen,Ph.D.Assistant ProfessorDepartment of Information ManagementFu Jen Catholic Universitypage 2TFT-LCD Research_GroupCopyright 2010 研究背景與動機nTFT-LCD

2、面板產業是由三大生產製程階段所組成，分別為列陣（Array）製程、組立（Cell）製程和模組（Module）製程，每階段製程由多個工廠所組成而形成稱為生生產產鏈結構鏈結構（production chain）的生產系統。n由於下列三大趨勢，而使的產能規劃問題變成了TFT-LCD產業日益重要：n多階層多階層、多世代與多廠區共存多世代與多廠區共存的生產鏈結構 n複雜的複雜的產產品階層結構品階層結構而造成了TFT-LCD面板可生產產品種類繁多且廣泛 n成長快速成長快速與劇烈變動劇烈變動的產產品需求品需求 n在這些有限產能供給、特殊生產結構與快速成長需求的特性下，TFT-LCD產業必須面臨了因供給與需求

3、不平衡而造成的產能規劃議題 n產產能分配決策能分配決策：決定每個生產廠區的最大化利潤產品族組合以及每個廠區對於每個產品的最適生產數量 n產產能擴充決策能擴充決策：決定哪種類型的產能（各廠區總產能或各廠區對各產品族產能）在哪些廠區需要擴充或是要設立新的生產廠區來增加產能、該類型的產能有多少量需要被擴充或增加以及要透過購買多少機台或是附屬資源來擴充產能 page 3TFT-LCD Research_GroupCopyright 2010 研究目的n本研究主要針對TFT-LCD產業的產能規劃問題進行探討，本研究的目的整理如下：n定位TFT-LCD產業之產能規劃問題，並從TFT-LCD產業的供給面特性

4、、需求面特性與供給與需求不平衡特性來分析、定義與分類該產能規劃問題。n建構確定性產能規劃問題之數學模式，包含了單階層多廠區產能規劃問題與多階層多廠區產能規劃問題，並提出有效率的陰影價格為基之分解陰影價格為基之分解演算法演算法(shadow-price based decomposition；SPBD)來求解該數學模式。n考量需求不確定環境下，建構隨機性產能規劃數學問題之數學模式，包含了單階層多廠區隨機產能規劃與多階層多廠區隨機產能規劃問題，並提出有效率的期望陰影價格為基之分解演算法期望陰影價格為基之分解演算法(expected shadow-price based decomposition；

5、ESPBD)來求解之。page 4TFT-LCD Research_GroupCopyright 2010 Capacity Planning in TFT-LCD Manufacturing page 5TFT-LCD Research_GroupCopyright 2010 Procurement Production Distribution Sales需求規劃（DemandPlanning）生產鏈規劃（Supply Network Planning）關鍵物料規劃（Critical Material Planning）訂單滿足規劃（OrderFulfillment）Module短期多廠區

6、最終日排程（Module Final Assembly Scheduling）Array、CF與Cell短期多廠區日排程（Array/CF/Cell Daily Scheduling）Array現場排程(ArrayOperationsScheduling)Cell現場排程（CellOperationsScheduling）Module現場排程（ModuleOperationsScheduling)ProductDemand ForecastDemand ForecastSupplyNetworkPlan MaterialPlanSupply Network Plan CellDailySche

7、dule Cell ATP ModuleATPAllocationPlanCustomerOrderArrayDailySchedule Cell DailySchedule Module FinalAssembly Schedule PUSHPULLMid-term(monthly)Short-term(daily)Shop floor(real time)產能規劃（Capacity Planning）Product FamilyDemand ForecastProduct DisaggregateTime DisaggregateTime DisaggregateResource Disa

8、ggregateHierarchical Planning Levels for TFT-LCD Industrypage 6TFT-LCD Research_GroupCopyright 2010 Planning Scenario of Capacity Planning Fab1.Fab2Fabn400,100 400,100 400,100 400,100 400,10 400,100 E175,100 175,100 175,100 175,100 175,10 175,100 D15,100 15,100 15,100 15,100 15,100 15,100 C50,100 50

9、,100 50,100 50,100 50,100 50,100 B150,100 150,100 150,100 150,100 150,10 150,100 A121110987Product GroupDemand(pieces/month)Unconstrained Demand Plan(forecast)Capacity Allocation&Expansion(Maximum total profit)Sales departmentMarginal profit262626292929E262626292929D252525282828AFab 2141414151515E77

10、7777C121212151515B151515181818AFab 112月月1110987Product GroupsiteProfit(US dollars/piece)89,493 86,575 89,493 86,050 88,950 88,950 Fab 276,720 74,240 76,720 74,240 76,720 74,000 Fab 1121110987siteCapacity of each site(sheet/month)8E8D6AFab 26E6C6B8AFab 1Value(pieces/sheet)Product GroupsiteOutputInput

11、Expansion Capability11234.70412089.82817066.88712456.23915719.23917895.493E40415.09539600.86834654.4939331.46236095.72333945.652D31548.11332297.20236847.86932545.05535521.93537500AFab 229839.80929839.80929839.80929839.80929839.80929839.809E2224.74722224.74722224.74722161.28582161.28582161.2858C8571.

12、13538571.13538571.13538575.53668575.53668575.5366B19696.09519696.09519696.09519737.53419737.53419737.534AFab 1121110987Product GroupsiteReleased Production Quantities(sheet)Capacity Allocation Plan(Release production quantities)0 0 0 0 0 0 E0 0 0 1 0 0 D0 0 0 0 0 0 AFab 20 0 0 0 0 0 E0 0 0 1 0 0 C0

13、0 0 0 0 0 B0 0 0 0 0 0 AFab 1121110987Product GroupsitePurchasing AmountCapacity Expansion Plan(Purchasing amount of new auxiliary tools)87028.50993652.646132206.9396461.115121729.79138582.7E313071.49306447.35267893.07303638.88278370.21261517.3D182909.65187252.72213636.58188592.09205842.51217305AFab

14、 2175100175100175100175100175100175100E12427.43812427.43812427.43812021.07212021.07212021.072C501005010050100501005010050100B150100150100150100150100150100150100AFab 1121110987Product GroupsiteSale Quantities(piece)Constrained Sales Plan(Sales quantities)Capacity on each siteCapacity of each site at

15、 certain site(sheet/month)Cutting Ratio37,500 37,500 37,500 37,500 37,500 37,500 E75,000 75,000 75,000 75,000 75,000 75,000 D37,500 37,500 37,500 37,500 37,500 37,500 AFab 233,000 33,000 33,000 33,000 33,000 33,000 E33,000 33,000 33,000 33,000 33,000 33,000 C33,000 33,000 33,000 33,000 33,000 33,000

16、 B33,000 33,000 33,000 33,000 33,000 33,000 AFab 1121110987Product GroupsiteCapacity of each site at certain site(sheet/month)12375,000 75,000 75,000 75,000 75,000 75,000 E75,000 75,000 75,000 75,000 75,000 75,000 D75,000 75,000 75,000 75,000 75,000 75,000 AFab 266,000 66,000 66,000 66,000 66,000 66

17、,000 E66,000 66,000 66,000 66,000 66,000 66,000 C66,000 66,000 66,000 66,000 66,000 66,000 B66,000 66,000 66,000 66,000 66,000 66,000 AFab 1121110987Product GroupsiteExpansion Upper Bound(sheet)Expansion Upper Bound1 1 0 0 1 Fab 21 0 1 1 1 Fab 1EDCBAsiteExpansion Capability4page 7TFT-LCD Research_Gr

18、oupCopyright 2010 產能規劃的角色n目的在於依據自於需求規劃（Demand Planning）模組所產生未來12個月未考量產能限制的產品族需求預測（Unconstrained demand forecast），決定未來12個月內各廠區（包括了Array廠區與Cell廠區）中還需購買多少資源（主資源與附屬資源）來擴充產能數量，並決定各廠區的最佳生產產品族組合與各產品族的生產數量 n產能規劃輸出結果的意義n提供建議給產能規劃人員來決定要在何時必須下採購單給資源設備廠商來購買擴充資源設備 n將產能被分配到生產產品族組合以及考慮產能限制下的可滿足需求預測數量（Constrained d

19、emand forecast）提供給需求規劃人員參考是否要重新調整需求預測來應付有限的產能資源 n經由產能規劃產生的產能擴充與生產分配資訊將會作為生產鏈規劃（Supply Network Planning）的輸入並驅動生產鏈規劃活動，生產鏈規劃就會依據既有的產能限制（初始產能加上從產能規劃獲得的擴充產能）下，決定Array、Cell與Module各階層每個廠區的6個月投入與產出生產量以及物料需求採購量 page 8TFT-LCD Research_GroupCopyright 2010 TFT-LCD產能規劃問題特性分析-供給面特性n生生產產鏈廠區結構特性鏈廠區結構特性n多階層生產環境 n多世

20、代技術與多廠區之共存 n多階層、多世代與多廠區生產鏈結構與關係 n多階層生產鏈間不同生產單位轉換 n多階層生產鏈的瓶頸漂移現象n生生產產鏈廠區與鏈廠區與產產品族關係特性品族關係特性n各世代廠區對不同產品族之經濟切割率n各廠區對不同產品族之生產能力限制n各廠區對不同產品族之潛在產能擴充能力(potential expansion capability)限制 n各廠區對不同產品族之生產良率 n各廠區對於生產不同產品族之生產良率為不同的，且每週期同一廠區加工不同產品族種類之良率也會有些微的變動 n對於新建立的生產廠區，由於廠區可能在剛設立（設立（ramp up）的階段，正在導入新製程技術以及調整機台

21、設定，公司將會慢慢調升其生產良率，造成每週期的良率呈上升的趨勢 n各廠區對不同產品族之生產變動成本 page 9TFT-LCD Research_GroupCopyright 2010 TFT-LCD產能規劃問題特性分析-供給面特性n生生產產鏈廠區供給鏈廠區供給產產能特性能特性 TFT-LCD產業的廠區供給產能可以分為兩種型態，一種為只考慮以廠為單位的各廠區總各廠區總產產能（能（Capacity of each site），另一種則必須考量以廠區加上產品族為單位的各廠區對各各廠區對各產產品族品族產產能（能（Capacity of each product group in certain si

22、te），此兩種型態的產能分別由不同設備（瓶頸機台或附屬資源）的數量來評定之。n各廠區總產能（Capacity of each site）限制與衡量 n各廠區對各產品族產能限制（Capacity of each product group in certain site）與衡量 n各廠區擴充產能方式 page 10TFT-LCD Research_GroupCopyright 2010 TFT-LCD產能規劃問題特性分析-需求面特性n產產品特性品特性 n複雜且多階層產品階層架構 n產產品需求特性品需求特性 n成長快速多期產品需求 n劇烈變動的不確定產品需求 n產產品價格特性品價格特性n多期產品平

23、均銷售價格（Average Sale Price；ASP）n劇烈變動的不確定產品平均銷售價格 nTFT-LCD產業不僅具有劇烈變動的不確定產品需求外，也具有劇烈變動的不確定產品平均銷售價格，此也是一項重要的因素影響著產能規劃結果無法準確的得到TFT-LCD穩健產能分配與擴充計劃 page 11TFT-LCD Research_GroupCopyright 2010 TFT-LCD產能規劃問題特性分析-供給與需求不平衡特性nTFT-LCD產業的產品生命週期短、需求成長快速以及產品價格劇烈波動，在加上新世代製程技術與廠區的建立需要花上長久的前置時間，因此市場常常會出現供過於求或供應求的況(供給與需

24、求不平衡的狀況)。n產能規劃的目的必須考慮整個規劃時程之供給產能與需求預測資訊，在規劃時程內，因此產產能規劃必須平衡此供給與需求不協調能規劃必須平衡此供給與需求不協調的狀況，將資源作最好的分配，使得最後規劃結果達到總利潤最佳化。n需求預測大於供給可用產能（供應求）n產產能分配（能分配（Capacity Allocation）n決定每個生產廠區的最大化利潤產品族組合以及每個廠區對於每個產品的最適生產數量 n產產能擴充（能擴充（Capacity Expansion）n決定哪種類型的產能（各廠區總產能或各廠區對各產品族產能）在哪些廠區需要擴充或是要設立新的生產廠區來增加產能，以及該類型的產能有多少量

25、需要被擴充或增加n產能擴充決策上，由於本研究所考慮的是透過購買某產品族副屬資源的方式來擴充各廠區對各產品族產能，並且還必須考量未滿足需求之產品族所剩餘之生命週期以及需求趨勢 n透過產能決策決定哪些產品族將要移轉到（移轉到（migration）哪些新的廠區進行生產n需求預測小於供給可用產能（供過於求）n產產能分配（能分配（Capacity Allocation）n決定每個生產廠區的最大化利潤以及最小化總成本的產品族組合以及每個廠區對於每個產品的最適生產數量。page 12TFT-LCD Research_GroupCopyright 2010 TFT-LCD產能規劃問題分類page 13TFT-

26、LCD Research_GroupCopyright 2010 Deterministic Multi-Site Capacity Planning Tzu-Li Chen,James T.Lin and Shu-Cherng Fang,“A Shadow-Price Based Heuristic for Capacity Planning of TFT-LCD Manufacturing”,Journal of Industrial and Management Optimization,Vol.6,No.1,pp.209-239,2010.page 14TFT-LCD Research

27、_GroupCopyright 2010 Problem definition of multi-site capacity planningnUnder the single-stage&multiple-site structure,multiple-product groups and multiple periods environments,each site with the specific generation can produce many different product group and each product group can be produced in m

28、any sites with different generation.nAssume demand forecast and sale price of each product group,capacity(supply)and cost information of each site for each product group in the future period are given and deterministic.nUnder given foregoing characteristics and constraints,a multi-site capacity plan

29、ning problem of the TFT-LCD industry consists of two main decisions to maximize total net profit:Capacity allocation decisionnThe profitable“product mix”of each site in each periodnThe best“production quantities”of each product group at each site in each period Capacity Expansion decisionnThe“purcha

30、sing amounts of the auxiliary tool(Mask)”for each product group at each site in each periodnThe“capacity expansion quantity”for each product group at each site in each periodnThrough outcomes of capacity expansion,a new product transfer plan that improves the flexibility configurations of multi-site

31、 structures is simultaneously identified.page 15TFT-LCD Research_GroupCopyright 2010 Specific characteristics of multi-site capacity planningOnly consider Array StageSingle-Stage&Multi-Site StructureRelease and output production balance constraint according to economic cutting ratioProduct group lev

32、el of the product hierarchy Limited flexibility configurationProduction capability constraint between each site and each product groupDifferent yield rate of each product group in each siteDeterministic demand and vary by periodsDeterministic price and vary by periodsAggregated demand forecast and s

33、ale priceGlobal capacity of each siteProduct-group-specific capacity at a certain siteOnly focus on purchasing the “mask”to expand capacity of each product group in certain siteCapacity consumption ratePotential expansion capability constraint Capacity expansion upper constraintpage 16TFT-LCD Resear

34、ch_GroupCopyright 2010 Multi-Site Capacity Planning Problem-AssumptionsnDemand quantities and sale prices of each product group are given and varied by period.nOnly consider Array process without Cell and Module processes.Since the Array process is the bottleneck,in a high-investment production envi

35、ronment,todays capacity planners focus majorly on solving the single-stage and multi-period capacity planning problem.nVariable production cost(including material cost,labor cost and other manufacturing cost)and holding cost depend on average unit cost.nCapacity expansion focuses on purchasing new a

36、uxiliary tools without adding new bottleneck machines or building new sites.nThe phase-out time of a product group can be estimated.nCalculation methods of capacity expansion cost adopt the“Straight-Line Method”(also called Linear Depreciation Method).nThe salvage value of each auxiliary tool is zer

37、o.page 17TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model-Notationpage 18TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model-Notationpage 19TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model -Objective FunctionnObjective

38、Function-maximize the total net profit Total RevenueTotal Variable Production CostTotal Inventory Holding CostTotal Capacity Loss CostTotal Expansion Costpage 20TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model -Objective FunctionnStraight-Line MethodnAdopt Straight-Line M

39、ethod to calculate the capacity expansion cost.estimated useful life of the auxiliary(Mask)using the number of time periods until the period T page 21TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model -ConstraintsConnector between capacity allocation and expansion problems!

40、page 22TFT-LCD Research_GroupCopyright 2010 Multi-Site Capacity Planning MILP Model -Constraintspage 23TFT-LCD Research_GroupCopyright 2010 Stochastic Multi-Site Capacity Planning under Demand UncertaintyLin,J.T.,Wu,C.H.,Chen,T.L.and Shih,S.H.,“A Stochastic Programming Model for Strategic Capacity P

41、lanning in Thin Film Transistor Liquid Crystal Display(TFT-LCD)Industry”,Computers and Operations Research,accepted.page 24TFT-LCD Research_GroupCopyright 2010 Problem definition of stochastic multi-site capacity planningnSince demand forecasts are usually inaccurate in TFT-LCD industry,traditional

42、deterministic capacity planning model is not reliable and no longer enough to tackle this problem.The stochastic multi-site capacity planning model is developed.nUnder the single-stage&multiple-site(f)structure,multiple-product groups(p)and multiple periods(t)environments,each site with the specific

43、 generation can produce many different product group and each product group can be produced in many sites with different generation.nWe use a scenario tree with discrete demand scenarios to represent the demand uncertainties and each scenario is associated with a given probability.Each scenario spec

44、ifics demand volumes of each product group over the planning horizon.nUnder the given TFT-LCD characteristics and demand uncertainty,the stochastic multi-site capacity planning problem addresses two-stage decisions to maximize the expected total profits.nIn the first stage,due to the long procuremen

45、t lead time of auxiliary tools,the capacity expansion decision(also called here-and-now decision)will determine the robust purchasing quantities of the auxiliary tool for each product group at each site before the actual demand is known.nIn the second stage,when a specific demand scenario is realize

46、d,capacity allocation decision(also called wait-and-see decision)will then generate a profitable product mix and production quantities for each site for each scenario.page 25TFT-LCD Research_GroupCopyright 2010 A scenario tree with discrete demand scenarios for representing the demand uncertainties

47、page 26TFT-LCD Research_GroupCopyright 2010 Problem definition of stochastic multi-site capacity planningpage 27TFT-LCD Research_GroupCopyright 2010 Modeling Demand Uncertainty and Scenario Generation nIn order to build the representative scenario tree with several discrete demand scenarios,a method

48、ology that combines the forecast techniques and scenario generation methods is proposed to approximate the stochastic demand process.page 28TFT-LCD Research_GroupCopyright 2010 Modeling Demand Uncertainty and Scenario Generation(1)-Time series forecast model (TFT-LCD case)nHistorical and predicted d

49、emand date of all products in the TFT-LCD manufacturing nThe mean demand,standard deviation and 95%confidence intervals of three products in each period page 29TFT-LCD Research_GroupCopyright 2010 Modeling Demand Uncertainty and Scenario Generation(2)-Demand scenario generation(TFT-LCD case)n100 sim

50、ulation scenarios(sample paths)of all products by the monte carlo sampling page 30TFT-LCD Research_GroupCopyright 2010 Modeling Demand Uncertainty and Scenario Generation(3)-Demand scenario reduction(TFT-LCD case)nDifferent demand scenario sizes of all products by the scenario reduction page 31TFT-L

51、CD Research_GroupCopyright 2010 Two-Stage Stochastic Programming Model of Multi-Site Capacity Planning-Notationpage 32TFT-LCD Research_GroupCopyright 2010 Two-Stage Stochastic Programming Model of Multi-Site Capacity Planning-Objective FunctionnObjective Function-maximize expected total net profitsE

52、xpected Total RevenueExpected Total Production CostExpected Total Inventory CostTotal Expansion CostFirst-Stage ObjectiveSecond-Stage Objectivepage 33TFT-LCD Research_GroupCopyright 2010 Two-Stage Stochastic Programming Model of Multi-Site Capacity Planning-ConstraintsnFirst-Stage Constraints(non-sc

53、enario related)page 34TFT-LCD Research_GroupCopyright 2010 Two-Stage Stochastic Programming Model of Multi-Site Capacity Planning-ConstraintsnSecond-Stage Constraints(scenario related)Connector between capacity allocation and expansion problems!page 35TFT-LCD Research_GroupCopyright 2010 Expected Sh

54、adow-Price based Decomposition-Scenario-dependent capacity allocation phase nWithout considering capacity expansion decision,our stochastic mixed integer programming model becomes the linear programming-based stochastic multi-site capacity allocation model(SMSCA)below.nSince the stochastic multi-sit

55、e capacity allocation problem is a large-scale linear programming model with the huge number of demand scenarios,we can decompose this whole model into a number of capacity allocation sub-models with individual demand scenario.page 36TFT-LCD Research_GroupCopyright 2010 Industry Practice and Model V

56、alidation-Numerical Study and Model Robustness nIn order to show the robustness of solution generated by two-stage stochastic programming,we conduct detailed numerical study to compare the solution robustness between the two-stage stochastic programming model(SP model)and the deterministic model usi

57、ng expected forecast demands(EV model).n A set of sample data includes two Array sites,five product groups,six months and three demand scenarios(low,medium and high demand)is collected from TFT-LCD industry partners.Robust auxiliary tool purchasing plan in SP model Auxiliary tool purchasing plan in

58、EV model page 37TFT-LCD Research_GroupCopyright 2010 Industry Practice and Model Validation-Numerical Study and Model Robustness nTo compare the SP model with the current industry practice(EV model),we use a stochastic measure,value of stochastic solution(VSS),to evaluate the performance of each mod

59、elnAccording to this result,using the SP model could gain 6.57%more profit than EV model under the given demand scenarios.SP represents the objective value of two-stage stochastic programming model.EEV represents the expected results of using solution of the expected-value deterministic model.page 3

60、8TFT-LCD Research_GroupCopyright 2010 Industry Practice and Model Validation-Robust Evaluation by out-of-sample Simulation nTo verify the effectiveness of stochastic model under normally distributed demand,we propose the simulation experimental framework to use out-of-sample simulation to study the

61、performance of EV and SP model in 150 randomly generated demand patterns.nThe total profit of each out-of-sample scenario under the fixed capacity expansion solutions of SP model and EV model is calculated and the profit distribution from the collection of the profit values of all out-of-sample scen

62、arios is formed.page 39TFT-LCD Research_GroupCopyright 2010 Industry Practice and Model Validation-Robust Evaluation by out-of-sample SimulationnThe mean,VaR&CVaR measure of the numerical example nThe distribution of the profit for the SP and EV model in the illustrative example*Improvement gap=(a b

63、)/b,where a is the value of SP model and b is the value of EV modelThe mean value of SP model is larger than EV model,so the solution of SP model provides higher expected profits in the face of the demand uncertainty.In terms of 95%VaR,90%VaR,95%CVaR and 90%CVaR,the objective function value of SP mo

64、del is also larger than that of the EV model.That means the worst return profits(VaR)and worst expected profits(CVaR)using the SP model are greater than that of EV model under 90%and 95%confidence level.page 40TFT-LCD Research_GroupCopyright 2010 Industry Practice and Model Validation-Robust evaluat

65、ion by other industrial case studies nThe mean,VaR&CVaR of SP and EV model under other industrial case studies It can be seen that the mean profit and financial risk measures(VaR and CVaR value)of SP model are always greater than that of EV model in all eight cases.page 41TFT-LCD Research_GroupCopyr

66、ight 2010 Future ResearchnStochastic Risk Programming with Risk MeasurenStochastic Dynamic Programming(Multi-Stage Stochastic Programming)nRobust Optimization with unknown Demand Distributionpage 42TFT-LCD Research_GroupCopyright 2010 Thank you for your attention!&Welcome your comments and questions!