CityLearn

Source

Core Claim

CityLearn is a configurable Gymnasium environment for building energy coordination, load shaping, and demand response. It is useful to this wiki because it exposes multivariate building-energy observations, exogenous context, continuous control inputs, and rewards or costs in a non-vision domain.

Dataset Notes

  • CityLearn models a virtual district composed of building energy models and distributed energy resources.
  • Flat CSV files provide action-agnostic observations such as calendar values, building loads, weather, carbon intensity, and electricity pricing.
  • Other observations are computed during simulation, including storage state of charge, net electricity consumption, device efficiencies, and dynamic indoor temperature in supported versions.
  • A schema controls the simulation window, episode splitting, time resolution, active observations, and active actions.

Action-Time-Series Notes

CityLearn has a clean control-input interface. The current documentation describes real-valued actions in [-1, 1] for storage charge/discharge or device power fractions. Named channels include cooling, heating, domestic-hot-water, and electrical storage, plus cooling and heating device controls.

For action-conditioned world-model work, the strongest use is to generate trajectories under fixed policies or candidate controllers:

building/district observations + weather/pricing/context + action_t
  -> next building/district observations + reward/cost

Gotchas

  • CityLearn is not a single immutable dataset payload; it is an environment and dataset schema.
  • Dynamics and available controls can depend on the schema and CityLearn version.
  • Source CSVs can be simulated, measured, or precomputed depending on how a dataset is constructed.
  • Repository code is MIT-licensed, but externally supplied building, weather, pricing, or carbon-intensity files should be checked before reuse.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Causal structure, counterfactuals, and controlpartially closesProvides explicit continuous control inputs for building energy storage and devices, with simulator transitions and reward/cost functions.Needs pinned benchmark protocols and versioned schemas before comparing general TSFM-style action-conditioned models.
Context interface: channel context and general contextpartially closesSchema and CSV files expose building metadata, weather, pricing, carbon intensity, and active observation/action definitions.Needs a reusable typed context schema across energy, industrial, and operations datasets.
Time representation and irregular event streamsadjacentUses explicit simulation time steps, episode ranges, and forecasted exogenous variables.Mostly regular simulation time; not a stress test for irregular event streams.
Benchmarks: what level of modeling is tested?adjacentStandardizes controller evaluation for building demand response and district coordination.Benchmark target is control performance, not a foundation time-series model suite with broad cross-domain transfer.