Stochastic Optimisation

Stochastic Optimisation#

GreenBubble supports two-stage stochastic capacity expansion: a single LP couples multiple annual scenarios (each with its own time series) so that investment decisions are shared across scenarios while dispatch is scenario-specific.

For the mathematical formulation see the PyPSA stochastic optimisation documentation.

—

Configuration#

Enable stochastic mode in config/config.yaml:

stochastic:
  stochastic: true
  scenarios:
    '2022': 0.10    # year: probability (must sum to 1)
    '2023': 0.30
    '2024': 0.30
    '2025': 0.30
  CO2_cost_s:       # CO₂ cost (€/t) per scenario
    '2022': 80
    '2023': 100
    '2024': 100
    '2025': 120
  CO2_cost_ref_year_s:  # CO₂ cost already embedded in energy prices
    '2022': 0
    '2023': 0
    '2024': 0
    '2025': 0
  EVPI: true        # also compute Expected Value of Perfect Information

Each scenario year must have preprocessed input data. Snakemake downloads and preprocesses all scenario years automatically before building the coupled network.

See stochastic for the full parameter reference.

—

Running#

No special flags are needed — just run Snakemake as usual:

snakemake -n      # verify all scenario years appear in the plan
snakemake --cores 4

The stochastic token STC is appended to the output folder name instead of DET, so stochastic and deterministic results never overwrite each other.

—

EVPI#

When EVPI: true, GreenBubble runs one additional deterministic solve per scenario (with perfect foresight for that year) and computes the Expected Value of Perfect Information:

EVPI = E[cost under perfect information] − cost of stochastic solution

EVPI is written to networks/evpi.yaml in the output folder. It is automatically disabled when stochastic: false.

—

Limitations#

  • MILP (committable links) is incompatible with stochastic mode — PyPSA’s stochastic multi-scenario LP requires a pure LP (no binary variables). Set committable: false for all technologies when using stochastic mode. Note: committable + extendable is supported in deterministic capacity expansion via PyPSA’s big-M formulation; see committable-extendable example.

  • Temporal resampling with stochastic mode issues a warning (scenarios are resampled independently — see Temporal Resolution).

—

See also#