GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-06
A building-control proposal may promise lower electricity costs while leaving an important question unanswered: what service changed to obtain the saving? A battery strategy, charging schedule or thermal-control policy needs to be assessed alongside the people and activities it supports. The architectural contribution is to make those service expectations explicit before one favourable energy result becomes the decision.
CityLearn is an open-source simulation environment for building energy coordination, demand response and reinforcement learning. Its documentation provides a common setting for comparing controllers under shared scenarios. It supports research into operational choices; a simulation result remains conditional on the configured buildings, systems and data. This article proposes a review method using documented capabilities rather than reporting an independently reproduced experiment.
Begin with a small scenario and write down what must remain comparable. Fix the buildings, weather period, occupancy assumptions, equipment capacities, tariffs and available observations. Specify the services that matter to the study, such as thermal conditions or vehicles receiving sufficient charge before departure. If an alternative changes a battery size as well as its controller, identify that as a combined equipment-and-control option rather than attributing the whole difference to AI.
Run or commission a conventional comparison alongside the proposed learning-based controller. State exactly what the reference does. The CityLearn scorecard guide distinguishes its conventional evaluation counterfactual from a separate business-as-usual rollout, and warns against mixing their ratios. A result described simply as a percentage improvement is incomplete unless the denominator and retained equipment are clear.
Read service measures beside energy and cost. A lower peak could be unhelpful if it coincides with unmet charging requirements; a thermal strategy needs its comfort-related indicators included. Preserve unavailable values rather than turning them into zero. Show absolute quantities with units as well as normalised comparisons so that the team can understand the scale of a difference and not just its direction.
Separate the controller's requested action from what the simulated system actually applies. The features documentation describes recording requested, limited and applied actions under electrical-service constraints. This distinction helps reveal a policy that repeatedly asks for unavailable capacity even when constraint handling prevents the request from taking effect. Such behaviour matters when discussing equipment sizing and the supervision required around a controller.
Review individual buildings and difficult periods before accepting the community average. A shared improvement can conceal a poor outcome for one participant or a short episode when service deteriorates. Identify the responsible assumptions and repeat comparisons under clearly defined variations. These are proposed robustness questions for the study, not a claim that any controller has already met them.
The handover should explain which strategy was compared, which service conditions were preserved, where constraints intervened and what remains uncertain. For architects and building operators, that is a more useful basis for an energy discussion than a single reward curve. Operational intelligence becomes relevant to design when its claimed benefit can be traced back to the spaces, systems and services that produced it.