Weather/outage-aware recourse

This comparison adds an assumed annual crew calendar to the rolling-recourse policy. Each year observes cost, hours, and capacity, and the execution calendar contains five fixed outage days plus independent 20% weather-loss probability per day in a 75-day fall window.

Three cases are compared:

  • calendar-naive: keep the resource-feasible plan and accept partial execution;

  • forecast-aware: plan against an independent synthetic forecast calendar, then execute under the actual calendar;

  • calendar-aware: defer unscheduled actions until every retained action fits the observed calendar.

The forecast is a synthetic draw from the same assumed distribution, not a forecast product. The observed-calendar case uses within-year calendar information before deferral. These are calendar policies, not stochastic dynamic programs.

Results

20 episodes, seed 7.

Case

Budget violation

Completed scope P50

Retained scope P50

Mean risk

Cost P50

Unscheduled h

Calendar deferrals

Calendar-naive

0.0%

88.4%

99.4%

0.0480

297,595

605.9

0.0

Forecast-aware

0.0%

86.2%

88.3%

0.0526

264,470

64.7

39.2

Calendar-aware

0.0%

87.6%

87.6%

0.0502

264,047

0.0

39.9

Forecast awareness reduces unscheduled work from 605.9 h to 64.7 h, but forecast error remains. Observed-calendar awareness eliminates unscheduled retained work but costs about 1.4 percentage points of mean scope versus forecast awareness and raises mean risk. Neither aware case lowers risk; both convert partial work into deferred scope.

590-segment scale test

A replicated 590-segment instance scales the workforce to 30 crews and 30 machines. The fixed plan is exact; annual re-solves use Fresh Solutions.

Case

Budget violation

Completed scope P50

Retained scope P50

Mean risk

Cost P50

Unscheduled h

Calendar-naive

0.0%

85.9%

99.6%

0.0617

2,942,470

7,377

Forecast-aware

0.0%

83.6%

86.4%

0.0713

2,539,520

681.5

Calendar-aware

0.0%

84.5%

84.5%

0.0667

2,536,614

0.0

Runtime: fixed exact 48.6 s; 20 weather episodes 220.8 s. Forecast error leaves 681.5 h unscheduled at scale, while observed-calendar awareness eliminates unscheduled retained work. The instance is replicated, not a real network, and the heuristic makes no global-optimality claim.

Limits

  • Forecast awareness uses a synthetic calendar draw, not a weather forecast.

  • Observed-calendar awareness uses within-year information before deferral.

  • Weather loss is independent per day and assumed, not measured.

  • The outage block is assumed, not an asset-owner schedule.

  • Whole actions are deferred when they cannot be completed, so the aware cases are conservative.

  • No safety, asset-owner priority, access, or field-calibration claim.