Академический Документы
Профессиональный Документы
Культура Документы
&Chemical
Engineering
Computers and Chemical Engineering 24 (2000) 329-335
www.elsevier.com/locate/compchemeng
Abstract
Model Predictive Control is a well established technique for the control of processes and plants. We present a similar concept
for planning and scheduling problems. There have mainly been two approaches to solve the planning and scheduling problems.
The first approach is to model the planning and scheduling as one monolithic problem and solve it for the entire horizon. Needless
to say, this approach requires an extensive computational effort and becomes impossible to solve in the case of large-scale
scheduling problems. The other approach is to hierarchically-decompose the problem into a planning level problem and a
scheduling level problem. This approach leads to tractable problems. Neither of these approaches provide the framework for
incorporating uncertainties in the processing time of batches, or random equipment breakdowns, or demand uncertainties in the
future. Furthermore, these approaches only provide ‘one snapshot’ of the planning problem and not a ‘walk through the timeline’.
Model predictive planning and scheduling provides a framework for studying dynamics. Model predictive planning and scheduling
requires a forecasting model and an optimization model. Both these models work in tandem in a simulation environment that
incorporates uncertainty. The similarity with the model predictive approach which is widely used in process-control is that in each
period, the forecasting model calculates the target inventory (controlled variable) in the future periods. These inventory levels
ensure desired customer service level while minimizing average inventory. The scheduling model then tries to achieve these target
inventory levels in the future periods by scheduling tasks (manipulated variables). 0 2000 Elsevier Science Ltd. All rights reserved.
0098-1354/00/$- see front matter 0 2000 Elsevier Science Ltd. All rights reserved.
PII: SOO98-1354(00)00469-5
330 S. Bose, J.F. Pekny /Computers and Chemical Engineering 24 (2000) 329-335
2. What are the appropriate ways of measuring fore- ulated variables). This is repeated till the end of the
cast accuracy, specifically, the accuracy of forecast- planning period. Refer to Fig. 1.
ing promotional demand?
To address the issues given above, one must capture 3. Coordination structure of the supply chain:
the dynamics of the supply chain by modeling each definitions
element of the supply chain (the supply unit and the
individual markets) and the interaction among these The coordination structure of a supply chain is
elements. The framework chosen for this study captures jointly determined by its decision rights structure
the forecasting, optimization and simulation aspects of (whether the plant or the distributors decide what to
the supply chain. One must also classify the coordina- produce and when) and its information structure (local
tion structure by some quantifiable parameters. The knowledge of promotions, distributor’s stock visibility,
rest of the paper explains these in detail and also etc.). The coordination structures can be classified in
presents the results obtained from the study so far. the following categories.
Industrial data has been used to obtain results and
draw conclusions to be sure that realistic assumptions 3.1. Centralized control
have been made about supply chain behavior.
The quantities to be supplied to each market are
centrally determined. The forecasts for the demands are
2. The model predictive approach based on past sales and recorded data. The forecasts of
the promotional part of the demand incorporate infor-
Here we present a model predictive approach for mation that is centrally available, and do not include
solving this problem. Model predictive planning and local knowledge of promotions. Thus, in a centralized
scheduling requires a forecasting model and an opti- control, even though most of the promotional events
mization model. Both these models work in tandem in are known, the actual quantity of the promotional
a simulation environment that incorporates uncertainty. demand and the exact week of occurrence have a lot of
The similarity with the model predictive approach uncertainty.
which is widely used in process-control is that in each
period, the forecasting model calculates the target in- 3.2. Decentralized control
ventory (controlled variable) in the future periods.
These inventory levels ensure desired customer service The individual markets place orders with the supply
level while minimizing average inventory. The schedul- unit based on local promotional knowledge. Each indi-
ing model then tries to achieve these target inventory vidual market is unaware of the capacity constraints in
levels in the future periods by scheduling tasks (manip- the factory, as well as, demands placed by the other
mvu?
S. Bose, J.F. Pekny / Computers and Chemical Engineering 24 (2000) 329-335 331
markets. Thus, even though the forecast quantity l Uncertainty in the actual quantity of the promo-
of promotional demands is relatively accurate, the tional demand.
markets do not provide enough lead-time for l Uncertainty in the actual week of occurrence of the
SUPPlY. promotional demand.
3.3. Distributed Control For various values of the above, different forecasts
can be generated. With each forecast, a parameter c1
The quantities are determined using both central and can be associated which is defined as the fraction of the
local promotional knowledge. There is information total promotional demand that is expected (known).
sharing and negotiation among the supply units and
Known Prom. Demand
individual markets.
’ = Total Prom Demand
The parameter CImeasures the overall efficiency of the
4. Forecasting-optimization-simulation framework business processes in the supply chain.
Table 1
Range of values for control parameters
Parameters Regimes
the user as ‘Promotional Criteria’), then that week is 5.2. Simulation module
marked as a promotional week. This is done for all the
weeks. Now the base demand estimated again to fit the The simulation module simulates production in the
predicted base demand to on& those weeks, which are factory for one week, at a time during which the
not marked as promotional. The sale for each week is schedule is frozen. The inventory levels at the market
compared with the predicted base demand and the warehouses are updated accounting for the actual de-
weeks, which meet the promotional criterion, are mands and allowing time for transportation of stock
marked and added to the previous set of promotional keeping units.
weeks. This iteration is performed until no more weeks Each of the three modules is invoked for each period
can be marked as promotional weeks. of the plan.
Table 2
Relation between control parameters and coordination structure
4.2. Centralized
6.3. De-centralized
level. Obviously, increasing the inventory safety stock- between 0 and 1. The u values were chosen to be
levels result in higher customer service level. The aim 0.2, 0.4, 0.6, 0.8 and 1.0, a total of five values. For
is to attain the desired customer service level of 99.7% each value of c1 the 27 combinations calculated above
with the minimum possible inventory. were simulated. That is a total of 27 times 5, i.e. 135
combinations. To determine the best possible customer
6.6. Computational effort service level in each run, six trials were made with
a different inventory level in each trial. That results
The number of simulations increases combinatori- in a total of 135 times 6, i.e. 810 combinations.
ally with the number of parameter and combinations Each run takes about 25 min for simulating 52 weeks
thereof. There are three control parameters (LT, Q on a Pentium II (400 MHz) processor. The total
and x), with three values each, hence, a total of 33 simulation time for all these runs is 20 250 min, which
times 3, i.e. 27 combinations. The parameter CI,which is about 14 days worth of computational time. Each
was introduced as a lumped parameter for measuring run is dependent on a random number seed, hence
the efficiency of business processes, can take any value for the same values of the parameters, several runs
should be simulated using a different random seed for
each run. It was found that the results of the runs do
not vary too much when a different random number
seed is used. This is because any peculiarities of a
random number are somewhat averaged over 52
weeks.
7. Results