SRC Project 879 Progress report Task 879.1: Intelligent Demand Aggregation and Forecasting Task Leader: Argon Chen Co-Investigators: Ruey-Shan Guo Shi-Chung Chang Students: Jakey Blue, Felix Chang, Ken Chen, Ziv Hsia, B.W. Hsie, Peggy Lin
Outline • Demand planning structure • Demand disaggregation methodologies • Demand grouping strategies • Fundamental study of demand planning approaches
Demand Planning Structure Demand Planning Structure Example � Two demands views: Time and Geography � Strategy: Top-down - Path 1: break down along Geography View first, then along Time View - Path 2: break down along Time View first, then along Geography View - Path 3: …… � Question: Which path? Demand Planning Structure: Optimal Demand Planning Path
Objective Objective � Objective: � Define performance metrics and a representation system for the multidimensional Demand Planning Structure. � Develop algorithms to search for the optimal Demand Planning Structure.
Performance Metrics Performance Metrics � Performance metrics depend on what the Demand Planning Structure is used for: � For forecast accuracy •Averaged CV as a measure Std.Dev of demand CV = = degree of fluctuation Mean of demand � For safety stock planning / auxiliary capacity planning •Sum of Standard Deviation as a measure
Representation System for Demand Representation System for Demand Planning Structure Planning Structure � View with Nested Attributes: e.g. time horizon (necessary), geo. view, etc.. Notation: View attribute • attribute Example: Time Year • Quarter • Month • Week Geography Continent • Country • City � View with Non-nested Attributes: e.g. product type Notation: View attribute × attribute Example: Product Generation × Function × Technology � View with Mixed Attributes Example: Product (Generation × Function × Technology) • PartNumber
Performance Metrics Computation Performance Metrics Computation Time Year • Quarter • Month • Week × Product Technology × Function CV CV CV CV Take average as an index!
Search Algorithm for Optimal DPS Search Algorithm for Optimal DPS � Heuristic 1: Top-down search � Heuristic 2: Bottom-up search � Heuristic 3: Middle-out search
Case Study Case Study � Demand Views � Nested-Attribute View: Time: Quarter, Month, Week Customer: Customer Region (CR), Customer Code (CC) � Mixed View: Product: Technology (T), Layers of Metal (L), Package (P); PartNum is nested to the combination of T, L, P
Testing Results Testing Results � Determine Demand Planning Structure: � Heuristic 2: Bottom-up search – Avg. CV as index Time Quarter • Month • Week x Customer CR • CC x Product (TxLxP) • PartNum Time Quarter • Month x Customer CR • CC x Product (TxLxP) • PartNum Time Quarter x Customer CR • CC x Product (TxLxP) • PartNum Time Quarter x Customer CR • CC x Product TxLxP Time Quarter x Customer CR x Product TxLxP Time Quarter x Customer CR x Product LxP Time Quarter x Customer CR x Product L Time Quarter x Customer CR x Product All Time Quarter x Customer All x Product All
Testing Results Testing Results � Determine Demand Planning Structure: � Heuristic 3: Middle-out search – Avg. CV as index Time Quarter x Customer All x Product All Time Quarter • Month x Customer All x Product All Time Quarter • Month • Week x Customer All x Product All Time Quarter • Month • Week x Customer All x Product L Time Quarter • Month • Week x Customer CR x Product L Time Quarter • Month • Week x Customer CR x Product LxP Time Quarter • Month • Week x Customer CR x Product TxLxP Time Quarter • Month • Week x Customer CR • CC x Product TxLxP Time Quarter • Month • Week x Customer CR • CC x Product (TxLxP) • PartNum
Testing Results Testing Results � Determine Demand Planning Structure: � Heuristic 1: Top-down search – Sum of Std. as index Time Quarter x Customer All x Product All Time Quarter • Month x Customer All x Product All Time Quarter • Month • Week x Customer All x Product All Time Quarter • Month • Week x Customer CR x Product All Time Quarter • Month • Week x Customer CR • CC x Product All Time Quarter • Month • Week x Customer CR • CC x Product P Time Quarter • Month • Week x Customer CR • CC x Product LxP Time Quarter • Month • Week x Customer CR • CC x Product TxLxP Time Quarter • Month • Week x Customer CR • CC x Product (TxLxP) • PartNum
Demand Disaggregation Demand Disaggregation A ggregating demand for better forecast → Disaggregate for detailed planning H ow to disaggregate? Total P1? P2? P3… P14? Asia Europe ….….. Africa
Impact of Product Life Cycle Impact of Product Life Cycle C haracteristics of semiconductor demands: - Fast-pace technology � fast-pace product life cycle - Highly substitutable A P roduct L ife C ycle A P roduct L ife C ycle Stage E xpand G row up M ature D ecline Stage E xpand G row up M ature D ecline C onventional disaggregation methods not sufficient to consider Effects of Product Life Cycle.
Conventional Disaggregation Methods Conventional Disaggregation Methods ● Method-A P i,m : proportion of product i at n = time bucket m ∑ P P n + i , n 1 i , t n : Total time period = t 1 ( Average the Proportion of previous “n” time buckets to determine the next one ) ● Method-B d i,m : demand of product i at n n ∑ ∑ d D time bucket m i , t t D m : Total demand at time bucket + = = = P t 1 t 1 m i , n 1 n n n : Total time period ( Average the demand of previous “n” time buckets to determine the proportion for the next time bucket)
Solutions - EWMA Statistic Solutions - EWMA Statistic ● To take into account the Product Life Cycle effect, “Exponentially Weighted Moving Average” (EWMA) can be used. ● Weight Calculating: Exponential Weight with r=0.02 0.05 i = week Weight 0.045 0.04 n = Total week number 0.035 0.03 0.025 0.02 0.015 = − − n i w r ( 1 r ) 0.01 0.005 i 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Time Exponential Weight with r=0.25 0.3 Weight For stable demand 0.25 0.2 r=0.01 0.15 0.1 For rapid-change 0.05 0 demand r=0.25 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Time
Decide the Proportion with EWMA Decide the Proportion with EWMA Note : In this case, we decide the Proportion with the 30 week data before it Sum of Weights = 1 Exponential Weight Time T=n-30 T=n-29 ………………….. T=n-2 T=n-1 T=n X History Data (Proportion or Demand) Time || T=n-30 T=n-29 ………………….. T=n-2 T=n-1 T=n+1 T=n P i,n+1
EWMA Disaggregation Method EWMA Disaggregation Method EWMA statistic is used in Method-A,Method-B. ● EWMA-A: Use EWMA of historical proportion to estimate the current proportion. n P i,m : proportion of product i at ∑ = ⋅ p w p time bucket m + i , n 1 t i , t W m : Exponential weights at time = t 1 bucket m ● EWMA-B: Use EWMA of historical demand to estimate the current demand and its proportion. n ∑ ⋅ w d d i,m : demand of product i at t i , t time bucket m = = p t 1 D m : Total demand at time bucket + i , n 1 n ∑ m ⋅ w D W m : Exponential weights at time t t bucket m = t 1
Test Data Description Test Data Description 1. Time horizon: 46-weeks semiconductor demand data. 2. Methods: conventional A, B; EWMA-A, EWMA-B 3. Historical data to determine proportions:30 weeks data Total P1 P2 P3… P14 Gen00 Gen01 ….….. Gen19
Test Methodology Description Test Methodology Description Performance Index : MSE ( Mean Square Error between forecast and real demand data ) • Different generations have different “r” values for best MSE performance. • The best “r” is used for each generation to estimate the proportion in Method-A and the demand in Method-B. For example... weight weight but r = 0.01 r = 0.25 Generation 04 Generation 07 (Demand is stable) (Demand is changing)
Preliminary Test Result Preliminary Test Result The result shows that : EWMA-B has the smallest MSE (best performance) MSE Comparison Total MSE Method-B 4,407,671 Method-A 5,572,988 EWMA-B 1,567,397 EWMA-A 3,086,232
Future Tasks Future Tasks 1. EWMA-B or EWMA-A? 2. Dynamically determine EWMA weight
Demand Grouping Strategies Demand Grouping Strategies � Demand grouping strategies to minimize inventory cost (see presentation material of liaison meeting on 7/25 and the attached technical paper “Aggregation Strategies to Minimize Inventory Cost under Uncertain Demand”) � Demand grouping strategies to minimize capacity requirement
Preparing Capacity Preparing Capacity capacity needed Capacity requirement (q)= OEE demand × processing time = OEE OEE : : Overall Equipment Efficiency Overall Equipment Efficiency OEE Capacity to be prepared = Average capacity requirement + Auxiliary Capacity = E(q) + α × St.Dev.(q) α is determined to satisfy a predetermined fill rate α is determined to satisfy a predetermined fill rate
Capacity Required Capacity Required Capacity Required Capacity to be prepared= E(q) + α × St.Dev.(q) � Observation from actual semiconductor demand data: Std.Dev of capacity demand = 0.5~1.5 Mean of capacity demand Auxiliary capacity is as important as average capacity � requirement!
Recommend
More recommend
