Industry: Project Management Product: ModelRisk Application: Resource Leveling
On a 200-activity urban transit build, the deterministic resource plan said the crawler-crane fleet would average 71% utilization across the 14-week station-shell phase — comfortably below the 85% red line. Then the team simulated the peak week. The chart below is the result, and it reframes the whole conversation: the peak-week utilization distribution has a mean of 104% and a median of 101%, with a 54% probability of breaching 100% capacity in at least one week of the phase. The reassuring 71% was never the number that mattered.
Resource leveling on a deterministic Gantt is a recipe for finding the bottleneck on the day it bites; the same problem under simulation surfaces it months in advance. A heavy civil contractor running roughly $3B of concurrent infrastructure work rebuilt its leveling workflow in ModelRisk. The goal was not to chase a single "optimal" calendar — there isn't one — but to find resource assignments that stayed feasible across the bulk of the joint duration-availability distribution.
Three uncertainties drive almost every leveling failure: task durations, crew/equipment availability on a given day, and the correlation between the two when weather or supply shocks hit shared inputs.
Task durations. Weather-sensitive tasks (excavation, deck pours, exterior steel) were modeled with PERT distributions calibrated to 6 years of internal as-built data: 10/20/45 days for typical station-box excavation, with the long right tail driven by groundwater finds. Prefabrication tasks, sheltered and well-instrumented, got tight Normal distributions (mean 14 d, σ = 1.6 d).
Labor availability. Daily crew turnout was modeled as Beta(8, 1.4) scaled to nominal headcount — a left-skewed shape with mean ≈ 0.85 and a thin tail of low-attendance days driven by storms and competing call-outs. Plain Normal here is wrong: it puts mass above 100% attendance, which is nonsensical.
Equipment availability. Each major asset (crawler crane, TBM, batch plant) used a two-state discrete model — operational (p = 0.93) or down for repair (p = 0.07) — with downtime duration LogNormal(μ = log(2.8), σ = 0.7) in days, mean ≈ 3.6 days, P90 ≈ 6.9 days.
Material lead time. LogNormal with mean 18 d, σ_log = 0.45 — the standard right-skewed shape of port-and-haul logistics.
Correlation matters: a storm front hits attendance, weather-sensitive task durations, and crane operability simultaneously. ModelRisk's correlation matrix carries a 0.55 rank correlation between "weather day" and the affected variables so the simulator doesn't pretend these shocks are independent.
A traditional spreadsheet level-loaded the station-shell phase to a peak crane utilization of 71%. The simulation says 71% is close to the mean weekly utilization — but the distribution of the peak week across the phase is much wider, centred on a mean of 104% (median 101%), with a P90 of 120% and an overall probability of at least one week breaching 100% across the phase of about 54%. The deterministic peak figure understated the simulated mean peak by 33 percentage points.
The point estimate hid a roughly even-odds chance of needing emergency crane rental at a $35k/day spot rate. Once overtime, spot rental and late-finish penalties are rolled together, the expected over-capacity cost worked out to about $4.4M for the phase, with a P90 of $11.2M.
A tornado over the P90 peak-utilization metric ranked the seven candidate levers.
Excavation duration tail dominates, contributing about 14 percentage points of the P10–P90 spread on peak utilization. Crane MTTR is second; weather-correlated attendance shocks third. This is the prioritized investment list: the dollars buying the most leveling robustness go to (a) a second geotechnical campaign to compress the excavation duration tail, (b) a service contract that cuts MTTR from 3.6 to 1.8 days, and (c) a contracted backup crew that activates above a defined attendance threshold.
The team compared the as-planned schedule against two alternatives — shift non-critical tasks two weeks right and add a second crawler crane for the peak quarter — by re-running the same simulation under each plan and comparing total phase cost (direct + overtime + spot rentals + late-finish penalties).
The as-planned baseline carries a mean phase cost of $20.4M, a P90 of $27.2M and a P99 of $49.4M — the long upper tail is the spot-rental and liquidated-damages exposure the deterministic plan never priced. Right-shifting the non-critical tasks trims this modestly (mean $19.2M, P90 $23.6M, P99 $44.5M) but leaves a substantial tail. The second crawler crane is the standout: by absorbing the peak it collapses the over-capacity cost almost entirely, dropping the mean to $17.2M (about $3.2M below baseline), the P90 to $17.1M (a $10.1M cut) and the P99 to $20.6M (a $28.7M cut). The roughly $1.1M of upfront crane capex is far smaller than the ~$4.4M expected over-capacity cost it removes, so the second-crane plan is not a mean-versus-tail trade-off at all — it dominates on both. For a contractor on liquidated-damages exposure that made it the obvious choice.
When the resource is a crane, "average utilization" is not the question — the question is the probability of a breach in any week. Monte Carlo simulation gives leveling that probability; a flat Gantt chart cannot.