Task 879.1: Intelligent Demand Aggregation and Forecasting Task - - PowerPoint PPT Presentation

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

SRC Project 879 Progress report

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

For safety stock planning / auxiliary capacity planning

• Sum of Standard Deviation as a measure

Std.Dev of demand Mean of demand

CV = = degree of fluctuation

Representation System for Demand Planning Structure Representation System for Demand Planning Structure View with Nested Attributes: e.g. time horizon (necessary),

• geo. view, etc..

Notation: Viewattribute•attribute Example: TimeYear•Quarter•Month•Week GeographyContinent•Country•City View with Non-nested Attributes: e.g. product type Notation: Viewattribute × attribute Example: ProductGeneration × Function × Technology View with Mixed Attributes Example: Product(Generation × Function × Technology) •PartNumber

Performance Metrics Computation Performance Metrics Computation

TimeYear•Quarter•Month•Week × ProductTechnology × 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

TimeQuarter •Month•Week x CustomerCR•CC x Product (TxLxP)•PartNum TimeQuarter•Month x CustomerCR•CC x Product (TxLxP)•PartNum TimeQuarter x CustomerCR•CC x Product(TxLxP)•PartNum TimeQuarter x CustomerCR•CC x ProductTxLxP TimeQuarter x CustomerCR x ProductTxLxP TimeQuarter x CustomerCR x ProductLxP TimeQuarter x CustomerCR x ProductL TimeQuarter x CustomerCR x ProductAll TimeQuarter x CustomerAll x ProductAll

Testing Results Testing Results

Determine Demand Planning Structure: Heuristic 3: Middle-out search – Avg. CV as index

TimeQuarter •Month •Week x CustomerAll x ProductAll TimeQuarter •Month •Week x CustomerAll x ProductL TimeQuarter •Month •Week x CustomerCR x ProductL TimeQuarter •Month •Week x CustomerCR x ProductLxP TimeQuarter •Month •Week x CustomerCR x ProductTxLxP TimeQuarter•Month •Week x CustomerCR•CC x Product TxLxP TimeQuarter •Month•Week x CustomerCR•CC x Product (TxLxP)•PartNum TimeQuarter •Month x CustomerAll x ProductAll TimeQuarter x CustomerAll x ProductAll

Testing Results Testing Results

Determine Demand Planning Structure: Heuristic 1: Top-down search – Sum of Std. as index

TimeQuarter x CustomerAll x ProductAll TimeQuarter •Month x CustomerAll x ProductAll TimeQuarter •Month •Week x CustomerAll x ProductAll TimeQuarter •Month •Week x CustomerCR x ProductAll TimeQuarter •Month •Week x CustomerCR •CC x ProductAll TimeQuarter •Month •Week x CustomerCR •CC x ProductP TimeQuarter •Month •Week x CustomerCR •CC x ProductLxP TimeQuarter•Month •Week x CustomerCR•CC x Product TxLxP TimeQuarter •Month•Week x CustomerCR•CC x Product (TxLxP)•PartNum

Demand Disaggregation Demand Disaggregation

Total

Asia

P1? P3… P2?

Europe ….….. Africa

P14?

Aggregating demand for better forecast → Disaggregate for

detailed planning

How to disaggregate?

Impact of Product Life Cycle Impact of Product Life Cycle Characteristics of semiconductor demands:

• Fast-pace technologyfast-pace product life cycle
• Highly substitutable

Conventional disaggregation methods not sufficient to consider

Effects of Product Life Cycle.

Stage Stage

Expand Expand Grow up Grow up Mature Mature Decline Decline

A Product Life Cycle A Product Life Cycle

Conventional Disaggregation Methods Conventional Disaggregation Methods

• Method-A

( Average the Proportion of previous “n” time buckets to determine the next one )

• Method-B

( Average the demand of previous “n” time buckets to determine the proportion for the next time bucket)

n P P

n t t i n i

∑ =

= + 1 , 1 ,

n D n d P

n t t n t t i n i

∑ ∑

= = + = 1 1 , 1 ,

P i,m : proportion of product i at time bucket m n : Total time period d i,m : demand of product i at time bucket m D m : Total demand at time bucket m n : Total time period

• To take into account the Product Life Cycle effect, “Exponentially

Weighted Moving Average” (EWMA) can be used.

• Weight Calculating:

i = week n = Total week number

Solutions - EWMA Statistic Solutions - EWMA Statistic

i n i

r r w

−

− = ) 1 (

Exponential Weight with r=0.02

0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 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

Weight

Exponential Weight with r=0.25

0.05 0.1 0.15 0.2 0.25 0.3 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

Weight

For rapid-change demand r=0.25 For stable demand r=0.01

Decide the Proportion with EWMA Decide the Proportion with EWMA

T=n+1 History Data (Proportion or Demand) Time T=n-30 T=n-29 T=n T=n-1 T=n-2 ………………….. Exponential Weight Time T=n-30 T=n-29 T=n T=n-1 T=n-2 …………………..

Sum of Weights = 1

X || P i,n+1

Note : In this case, we decide the Proportion with the 30 week data before it

EWMA statistic is used in Method-A,Method-B.

• EWMA-A: Use EWMA of historical proportion to estimate the

current proportion.

• EWMA-B: Use EWMA of historical demand to estimate the

current demand and its proportion.

EWMA Disaggregation Method EWMA Disaggregation Method

∑

= +

⋅ =

n t t i t n i

p w p

1 , 1 ,

∑ ∑

= = +

⋅ ⋅ =

n t t t n t t i t n i

D w d w p

1 1 , 1 ,

d i,m : demand of product i at time bucket m D m : Total demand at time bucket m W m : Exponential weights at time bucket m P i,m : proportion of product i at time bucket m W m : Exponential weights at time bucket m

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

Gen00

P1 P3… P2

Gen01 ….….. Gen19

P14

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...

Generation 04 (Demand is stable) r = 0.01 weight

but

Generation 07 (Demand is changing) r = 0.25 weight

The result shows that :

EWMA-B has the smallest MSE (best performance)

Preliminary Test Result Preliminary Test Result

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 requirement (q)= =

OEE OEE : : Overall Equipment Efficiency Overall Equipment Efficiency

Capacity to be prepared = Average capacity requirement + Auxiliary Capacity = E(q) + α × St.Dev.(q)

α α is determined to satisfy a predetermined fill rate is determined to satisfy a predetermined fill rate

capacity needed OEE demand × processing time OEE

Capacity Required Capacity Required Capacity Required

Capacity to be prepared= E(q) + α × St.Dev.(q)

• Observation from actual semiconductor demand data:
• Auxiliary capacity is as important as average capacity

requirement! Std.Dev of capacity demand Mean of capacity demand

= 0.5~1.5

Demand Grouping for Production Demand Grouping for Production

3 groups 2 groups 2 groups 2 groups 1 group

A B C A B C C B A A C B B C A

Allocate machines for product demands

• Group product demands for machines

e.g. 3 demand sources: A, B & C; 5 combinations

Capacity requirement = 1.capacity needed (=demand × processing time)

Grouping product demand → fluctuation of capacity demand ↓ → auxiliary capacity ↓

Impacts of Aggregation on Capacity Required Impacts of Aggregation on Capacity Required

Aggregate +

Time Time

Before aggregation

Time

After aggregation

Capacity needed OEE

Capacity requirement =

• 2. Overall Equipment Efficiency

Time Processing Actual Time Processing l Theoretica yield Time Total Time Production × × =

demand × processing time OEE

Impacts of Aggregation on Overall Equipment Efficiency Impacts of Aggregation on Overall Equipment Efficiency

After grouping products for the same machine

• machine flexibility , and the production time
• yield
• overhead time , and actual processing time

OEE may either improve or deteriorate after grouping.

Objective Objective

Objective: Minimum total capacity required (= Average capacity

requirement + Auxiliary Capacity)

Subject to

• Pre-determined demand fill rate
• Production grouping constraint
• Production scheduling flexibility
• Production yield
• Production changeover and overhead

Fundamental Study of Demand Planning Approaches Fundamental Study of Demand Planning Approaches

• Aggregation and forecasting approaches

(see presentation material of liaison meeting on 7/25 and the attached technical paper “Aggregation and forecasting approaches for two interrelated demand sources”)

• Aggregation, forecasting and disaggregation approaches

Aggregation, Forecasting and Disaggreation Aggregation, Forecasting and Disaggreation

• Aggregated demand is known for better forecast
• Disaggregated demand is needed for detailed planning
• Question:
• Should we forecast aggregated demand and then

disaggregate for detailed planning? OR

• Should we forecast and plan with the detailed demand
• nly?

Time Series Model and Corresponding Approaches Time Series Model and Corresponding Approaches

• Study Vehicle: VAR(1) demand model
• There are 5 possible aggregation, forecasting and

disaggregation approaches

t t t t t t t t

a X X X a X X X

2 1 2 22 1 1 21 2 1 1 2 12 1 1 11 1

+ + = + + =

− − − −

ϕ ϕ ϕ ϕ

with autocorrelation & cross correlation between them

Aggregation, Forecasting and Disaggregation Approaches Aggregation, Forecasting and Disaggregation Approaches

• Approach 1: No aggregation; No statistical forecasting.
• Approach 2: Aggregation & disaggregation; No statistical

forecasting.

• Approach 3: No aggregation; Univariate forecasting.
• Approach 4: Aggregation; Univariate forecasting;

Disaggregation;.

• Approach 5: No aggregation; Multivariate forecasting.
• Extensions: weighted aggregation in approaches 2 and 4.

Objective Objective

• Use a bivariate time series model as the study vehicle to study

the effects of aggregation, forecasting and disaggregation

• Study and compare different aggregation, forecasting and

disaggregation approaches

• Investigate the possibility of weighted aggregation

methodologies and corresponding forecasting approaches.