One net reading, multiple physical explanations
A low grid-import reading can mean low demand, high solar production, battery discharge, or a combination. Low-frequency aggregate telemetry also loses many switching transients used by appliance-disaggregation methods. The research problem is therefore identifiability: whether the observations contain enough information to justify an attribution, rather than simply produce a plausible decomposition.
Keep measured channels outside the unknown
The FY2026 programme established a multi-source telemetry substrate and confirmed that solar and battery channels can be separated at source. Those directly observed flows should be preserved before attempting load attribution. Absorbing them into an undifferentiated aggregate discards information the system already has. The resulting research direction is channel separation followed by attribution, evaluated against retained site telemetry.
Attribution must carry an error model
Evaluation requires device- or circuit-level estimates compared with sub-metered reference measurements, cost reconciliation against the bill, the fraction of load explained and calibrated confidence. Stability across homes with and without solar and storage is a separate test. The FY2026 report describes the telemetry foundation and research framing; attribution accuracy and confidence calibration were still being consolidated. No general disaggregation accuracy or savings percentage is asserted here.
Battery dispatch is constrained optimisation
An executable dispatch schedule must account for load forecasts, photovoltaic forecasts, time-of-use prices, battery reserve and a hard grid-import ceiling. A demand charge adds a path-dependent term: the worst fifteen-minute sliding average in a billing month can dominate a sequence of individually sensible interval decisions. A short-horizon controller therefore has to reason about both immediate energy cost and the accumulated peak exposure.
A predictor must beat a working baseline
The recorded deterministic layer uses predefined shaving windows, reserve state of charge and a hard floor. It provides a functioning comparison point for forecast-driven dispatch. Back-testing must include stale telemetry, forecast drift and actuation latency; a schedule that is better under perfect information may fail once those conditions are represented. The predictive core remains a research direction at the report boundary, distinct from the implemented deterministic controls.
Research outcome: a more precise question
The platform’s value is not measured by the presence of a predictor alone. It depends on whether decisions remain explainable, physically executable and better than a reliable alternative under the same constraints. Separating acquisition, attribution and dispatch makes each claim independently testable.