CityLearn

Summary

CityLearn is a Gymnasium environment and dataset schema for reinforcement-learning and control experiments in building energy coordination, load shaping, and demand response.

Official Artifacts

Dataset Shape

CityLearn combines virtual district/building models, flat time-series data files, schema-defined observation/action spaces, and runtime simulation. It can generate trajectories with observations, continuous control inputs, rewards/costs, and next observations.

Role In The Wiki

CityLearn belongs in the non-vision action-conditioned dataset bucket. It is especially useful as an energy-management simulator for testing how models handle exogenous context, building state, storage/device controls, and delayed effects.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in citylearn-2020 rather than duplicating verdict rows here.

At the entity level, CityLearn is a clean simulator-backed control environment rather than a passive forecasting dataset. Its main caveat is artifact/version pinning: experiments should record schema, package version, action set, reward/cost function, and source-data provenance.