Scenario-Based Capacity Simulation

Planners use scenario-based capacity simulation when the baseline S&OP plan is no longer a safe assumption. A tariff, an unplanned line outage, a delayed inbound lot, or a demand spike on a constrained family can make the published mix, load, and inventory trajectory infeasible within days. The job of the model is not to pick a winner. It is to generate comparable recovery paths, score them against the same constraints, and leave the choice of which path to act on with the planner.

Treat this as a decision-support loop, not an execution engine. Simulation and generative planning tools can propose overtime, alternate routing, safety-stock draws, expedites, or mix shifts. They must not auto-commit capacity calendars, purchase orders, or production orders. The planner selects the scenario, documents the trade-off, and only then hands approved actions to the systems of record.

A useful capacity scenario is a closed system: demand, supply, and the resources that convert one into the other. If any of those three is a slogan instead of a constraint, the ranking will look precise and still be wrong.

Demand needs a time-phased picture at the level you actually load the plant: family, work center, or SKU, depending on how you freeze the MPS. Pair this with Multi-Signal Demand Forecasting when the shock is a demand event rather than a supply event. A forecast revision, a channel fill-in, or a promotional pull-forward should enter the simulator as a volume and mix change, not as a single percentage on the total plant.

Supply needs more than on-hand inventory. Include open production, inbound POs with promised and worst-case dates, qualified alternates, and the lead-time distributions you already use for risk. Supplier Lead-Time Risk Scoring is the usual feed for supplier-failure and delay scenarios. A “supplier down” case that ignores qualification status, min order quantities, and inbound quality hold time will recommend a recovery that purchasing cannot place.

Capacity needs calendars, rates, changeover logic, and the bottlenecks you already know. Capacity Bottleneck Identification should tell you which resources to model in detail and which can stay as simple throughput limits. Over-modeling every helper resource slows the run without changing the ranking. Under-modeling the constraint that actually sets the weekly mix produces optimistic recoveries that fail on the shop floor.

Write each shock as an explicit event with a start, a duration or probability window, and an affected object (machine, supplier, lane, region, or demand stream). Tariff cases should hit cost and, where relevant, landed-time or origin mix, not only a margin comment. Downtime cases should hit available hours and, if true in your plant, the queue and WIP that cannot move to an alternate. Supplier-failure cases should hit receipts, not “risk score” as a unit of measure.

Once the shocked state is loaded, the model searches for feasible recoveries inside the planner’s allowed levers. Typical levers in manufacturing S&OP are finite capacity (overtime, extra shifts, weekend windows), alternate work centers or contract manufacturers, mix and allocation changes, inventory draws and postponement, expedited inbound, and delayed or split customer commitments where commercial policy allows it.

Keep the search bounded. An unbounded optimizer will invent recoveries you would never approve: unapproved suppliers, rates the line has never run, or overtime that labor agreements forbid. Encode those as hard constraints or as infeasible flags, not as soft penalties the ranker can bury.

Rank on a small, stable scorecard that S&OP already debates. Service (fill, OTIF, or late volume by priority class), inventory and cash (peak and ending), cost to recover (expedite, overtime, duty, scrap, changeover), and load feasibility (hours over demonstrated capacity, bottleneck utilization, freeze-fence violations) are usually enough. Add a “plan stability” measure: how much of the already-released horizon you would have to rip up. A recovery that saves a few late units by rescheduling every firm order is often worse than a slightly later recovery that protects the freeze.

Present results as a shortlist, not a single “optimal” plan. Three to seven named scenarios is enough for a weekly S&OP cycle: do nothing (absorb in inventory and backlog), recover with internal capacity only, recover with inbound expedite, recover with mix/allocation, and a combined case if those levers interact. Name them in planner language (“Line 3 down 10 days, recover with Plant B overflow + overtime”) so the meeting can argue the choice, not the filename.

Platforms that already do this class of work include Kinaxis and o9 Solutions for concurrent planning and scenario comparison in the S&OP stack, and AnyLogic when the constraint logic is discrete, stochastic, or layout-specific (queues, changeovers, material handling) and a spreadsheet or LP is a poor fit. Use the tool that matches how your plant actually fails. A network LP is the wrong engine for a changeover-dominated bottleneck. A detailed discrete-event model is wasted if the decision is purely “which DC ships Europe next month.”

Start from the published unconstrained or constrained baseline, not from last week’s sandbox. Snapshot demand, supply, and calendars so every scenario in the pack is comparable. If two planners load different inventory books, the ranking is theater.

Define the shock with operations and procurement in the room, or you will simulate the wrong outage. “Machine down” is not a scenario until you know whether WIP is stranded, whether a sister line is qualified, and whether maintenance can give a return-to-service window. “Tariff” is not a scenario until you know which origins, which duty rates, and whether you may re-source inside the planning horizon.

Run the do-nothing case first. If service and inventory stay inside policy, you may not need a recovery at all. Many shocks look dramatic in a slide and are absorbable in safety stock and a short backlog. Ranking recoveries without a do-nothing baseline hides that fact.

Then run the bounded recoveries. Check that each one respects freeze fences, qualified sources, and labor rules. Reject any result that requires a master-data change you have not approved (new routing, new supplier, new rate). Put those in a separate “if we qualify X” pack so the commercial or engineering decision is visible.

In the S&OP or supply meeting, the planner presents the shortlist with the same scorecard columns, states the recommended scenario, and records the reason. The output of the meeting is a chosen scenario identifier plus a list of human-approved actions: overtime request, alternate plant load, expedite PO, allocation hold. Those actions are entered by the people who own the transactions. The simulator does not write them.

After execution starts, keep the chosen scenario as the watch-point. If the outage lasts longer than the assumed window, or the expedite misses, re-rank. Do not leave a stale “winning” scenario in the deck after the facts have moved.

Inputs you should refuse to run without: item-location inventory that matches the ERP snapshot, a capacity calendar with holidays and already-approved overtime, demonstrated or standards-based rates on the constrained resources, open orders with dates, and a demand signal at the right aggregation. Missing any of those produces a ranking that cannot survive the first challenge from the plant manager.

Outputs you should insist on: time-phased load on the constraint, projected on-hand and backlog by priority, a list of violated constraints (if any), and an action list that maps to real transactions. A heat map without an action list is not a plan. An action list that says “increase capacity 12%” without saying which calendar, which week, and which approval is also not a plan.

Watch for ranking artifacts. If two scenarios differ only because the solver used a different unmet-demand penalty, the score is not a business comparison. If inventory looks better because the model delayed receipts you already paid for, the cash picture is fiction. If service looks perfect because the model silently dropped low-priority demand, say so in the meeting. The ranker should expose substitutions and unmet demand, not hide them in an objective function.

Stochastic or generative variants (sampled downtime, sampled lead times, sampled demand error) are useful when the shock itself is uncertain. Report ranges or percentiles on service and cost, and still pick a scenario to act on. A distribution is not a decision. The planner still chooses one recovery path for the next cycle, with a trigger to switch if a watched signal (return-to-service date, inbound ASN, booking pace) crosses a line.