Skip to content

Benchmarks overview

PowerZoo's public benchmark set is organised around five agent-centric task suites. Each suite targets a different RL research question and uses a different underlying environment, agent structure, action space and constraint regime; any two suites therefore differ on at least four of these dimensions.

The five suites

Suite Pillar Underlying env Agents Steps × Δt RL paradigm
TSO Security dispatch Transmission Case5 / Case118 1 (UC) / 54 (opf_118) 48 × 30 min (336 × 30 min for 7d) Safe RL with mixed discrete-continuous action
DSO DER control Distribution Case33bw + 6× FlexLoad + Ausgrid 1 48 × 30 min Non-stationary RL (loss minimisation)
DERs DER control Distribution Case33bw / Case118zh + heterogeneous DERs 3 / 5 / 12 48 × 30 min (168 for EV) Scalable safe MARL
DC microgrid DER control Self-contained DC microgrid 1 288 × 5 min Multi-objective robust RL
GenCos Market lite Transmission Case5 + market clearing 5 48 × 30 min Competitive MARL

Each suite targets a different RL paradigm, but all five share the same make_task_env interface and the same reward / cost split. No single algorithm performs best on all five suites; this is by design.

flowchart LR
    SD["Security Dispatch"] --> TSO[TSO]
    DERC["Intertemporal DER Control"] --> DSO[DSO]
    DERC --> DERS[DERs]
    DERC --> DCM[DC microgrid]
    MKT["Market Lite"] --> GC[GenCos]

Public PowerZoo task names per suite

Suite PowerZoo task(s)
TSO marl_uc (UC, Case5), opf_118, opf_118_7d (large-scale ED, Case118)
DSO make_dso_env(...) factory (assembled outside the standard registry)
DERs marl_der_arbitrage (Case33bw, 3 batteries), marl_ders_benchmark (Case118zh, 12 heterogeneous DERs), marl_ev_v2g (5 EVs)
DC microgrid dc_microgrid, dc_microgrid_safe
GenCos gencos_bidding

Smaller starter tasks sit outside the main set but remain in the public API: battery_arbitrage, marl_opf (5-bus MARL ED), dc_scheduling. They are the recommended first targets for unit testing and quick iteration; see Examples.

How to read a benchmark page

Every per-suite page follows the same template:

  1. Design intent and research question — what RL difficulty this suite is built for.
  2. Physical setup — case, resources, constraints.
  3. Agent design — observation, action, reward, cost.
  4. Variants — the in-suite difficulty ladder.
  5. Splits — train / val / test (or train / iid / OOD for DSO).
  6. Baselines — random, rule-based, and any oracle reference.
  7. Metrics — what to report.
  8. Code recipes — make_task_env(name) + a minimal training stub.

Difficulty axes across the five suites

The suites together span five orthogonal difficulty axes. You can cite this matrix in a paper to argue that the benchmark is diverse:

Axis Easiest → Hardest
Information structure TSO (centralised full info) DSO (single-agent non-stationary) → DERs (Dec-POMDP) GenCos (private-info competition)
Action complexity DSO (continuous) DC (multi-dim continuous) → DERs (per-agent joint) → TSO (mixed) GenCos (multi-agent strategic)
Safety regime GenCos (profit-driven) DSO (operational quality) → DC (SLA + thermal + power balance) DERs (hard voltage) → TSO (full safety)
Exogenous uncertainty TSO (intra-day variation) DSO (drift) → DERs (renewable) DC (workload + solar + thermal) → GenCos (+ adaptive opponents)
Objective complexity GenCos (profit) DSO (loss + serve) → TSO (cost + safety) DERs (safety + fairness) → DC (3-vector + constraints)

Each axis has a monotone ordering across the five suites, so picking different suites gives different RL difficulty rather than the same problem under different names.

What the public-API helpers report

from powerzoo.tasks import PUBLIC_TASKS, list_public_tasks, get_public_task_catalog

print(PUBLIC_TASKS)
print(list_public_tasks())
catalog = get_public_task_catalog()
print(catalog[0]['task_id'], catalog[0]['default_episode_horizon_steps'])

PUBLIC_TASKS includes the five main suites plus the smaller starter tasks. make_dso_env(...) is the only main task absent from PUBLIC_TASKS: it uses its own factory instead of the standard registry.

See also