Repository navigation
CityLearn 3.3.0
What's new
- Battery and EV physics. Activation respects the configured SOC below the discharge reserve. EVs retain their identity and evolve through charger removal, re-entry and handover. Accounting includes standby losses, mobility and explicit arrival energy. CSV and Parquet exports report the charger that actually served the EV.
- BAU and KPIs. Native BAU uses matching equipment, PV, actions and outages, including the same stochastic outages. Cost and net-exchange ratios, market comparisons and portfolio aggregation are corrected. Repeated evaluation retains only the latest BAU reference.
- Compatibility and robustness. Support NumPy 1 and 2 and legacy occupant models. Resume SAC and MARLISA training from 3.2.0 checkpoints. Keep EV session IDs, publish complete dataset caches atomically and make experiment evidence portable across checkouts.
Public dataset files and export layouts are preserved. If your scenario uses the corrected paths, rerun both the controller and its reference for comparisons. See the migration guide.
Validation
The integrated runtime passed the CI matrix on Python 3.9, 3.10 and 3.12, covering NumPy 1 and 2, runtime audits and installed-wheel exports. The battery and EV regression tests cover one-hour and 15-minute steps.
Credits
Thanks to Mark Colley (@M-Colley) for the cross-platform, BAU, NumPy, KPI, EV arrival, agent and dataset-cache fixes in #170–#175. Tiago Fonseca (@calofonseca) and the CityLearn maintainers reviewed, validated and integrated those changes, added matched stochastic outages, completed the EV session-ID and MARLISA fixes, and developed the battery reserve and EV lifecycle corrections in #179 and #180.