Scenarios and Probabilities#

The noise is the uncertain quantity revealed at each stage: the inflow in ClearLake. SDDP works with a finite set of possible noise outcomes per stage, the scenarios, each with a probability. You describe them with set_noise().

Describing the noise#

scenario_data = np.array([
    [ 50.0, 150.0, 350.0],   # jan
    [ 50.0, 150.0, 350.0],   # feb
    [-50.0, 100.0, 250.0],   # mar
    [-50.0, 100.0, 250.0],   # apr
])

sddp.set_noise(parameter=precip, scenario_data=scenario_data)

set_noise() takes:

parameter: the GAMSPy Parameter that carries the noise into your equations. You declare it empty; before each per-scenario solve the engine overwrites it with the value for the scenario being solved.
scenario_data: a 2-D array of shape (n_stages, n_scenarios): one row per stage, one column per scenario, holding the noise value for that combination. The ClearLake array above gives three inflow outcomes for each of the four months.

Probabilities#

By default the scenarios are equally likely. Pass probabilities to weight them, a 1-D array of length n_scenarios that is non-negative and sums to 1:

sddp.set_noise(
    parameter=precip,
    scenario_data=scenario_data,
    probabilities=[0.25, 0.50, 0.25],
)

The same probabilities apply at every stage. They weight the dual aggregation in the backward pass and the path sampling in the forward pass, so they shape both the cuts and the lower bound. Weighting scenarios unevenly is also the substrate for risk-averse training.

Note

set_noise() validates the inputs: scenario_data must be 2-D with one row per stage and at least one scenario column, and probabilities (when given) must be 1-D, match the number of scenarios, be non-negative, and sum to 1 within 1e-9. Omitting probabilities is equivalent to a uniform 1 / n_scenarios for every scenario.