cobre-sddp

Stochastic Dual Dynamic Programming (SDDP) algorithm for the Cobre power systems ecosystem.

Implements the SDDP algorithm (Pereira & Pinto, 1991) for long-term hydrothermal dispatch and energy planning. The crate covers the full solve cycle: forward pass scenario simulation, backward pass Benders cut generation, cut management with cut selection (Level-1, LML1, dominated-cut pruning, and Dynamic Cut Selection (DCS)), CVaR risk measures, convergence monitoring, policy warm-start and resume from checkpoint, and annual discount rate support. Designed for hybrid MPI + thread-level parallelism via ferrompi.

When to Use

Use cobre-sddp when you need programmatic access to the SDDP algorithm — embedding it in a custom orchestration layer, running parameter sweeps, or integrating the solver into a larger application. For single-study command-line use, prefer cobre-cli, which wraps this crate.

Key Types

• TrainingConfig — algorithm parameters: iteration budget, forward scenario count, checkpoint cadence, warm-start cut count, and cut selection strategy

• TrainingContext — runtime state shared across all iterations of the training loop (cut pools, workspaces, convergence monitor)

• CutPool — pre-allocated storage for Benders cuts with active/inactive bookkeeping

• CutSelectionStrategy — enum controlling cut pool pruning:

  • Level1 — deactivates cuts below tie_tolerance of the per-state max at every visited state
  • Lml1 — deactivates cuts that are not the oldest eligible within tie_tolerance at any visited state
  • Dominated — geometric dominance at visited states, using threshold as the tolerance
  • Dynamic — lazy incremental scheme (DCS): adds at most nadic cuts per inner re-solve round (the inner loop repeats up to max_inner_iterations rounds per backward solve) that violate the current LP solution by more than epsilon_viol (the CutSelectionStrategy::Dynamic API field; the config.json key is violation_tolerance, which falls back to tie_tolerance when absent); never deactivates cuts from the pool

  Both Level1 and Lml1 use tie_tolerance (default 1e-10) to control how closely a cut must approach the per-state maximum to be retained. The memory_window config field is deprecated and silently ignored; use tie_tolerance instead.

• SimulationConfig — parameters for the post-training simulation run

• ConvergenceMonitor — tracks lower bound, statistical upper bound, and gap closure across iterations

Links

Status

Alpha — API is functional but not yet stable. See the main repository for the current release.

License

Apache-2.0