Follow the filing velocity into an unexpected place: the grid-control CPC class is starting to fill with machine-learning methods. Two June 2026 grants, issued a week apart, make the trend concrete — and they sit in H02J, the class for power distribution and management, not in the G06N machine-learning class where you'd expect AI patents to land.
The first, US12651906B2, "Deep reinforcement learning agent for demand response in home energy management systems" (Qatar Foundation, issued June 9, 2026), classifies under H02J 3/17 — arrangements for reducing or controlling power demand. A reinforcement-learning agent decides when household loads shed or shift in response to grid conditions. The AI is the controller, and the CPC tag places it as a grid-control element.
The second, US12646943B2, "ML-optimized VPP controller for battery powered EV charging networks" (Banpu Innovation & Ventures, issued June 2, 2026), carries H02J 3/004 (regulation by means of converters) plus B60L 53/63 (EV charging control). Here a learning controller orchestrates a fleet of EV chargers and batteries as a single virtual power plant — aggregating distributed assets into one dispatchable resource.
Counts tell a strategy, even at small N. Two grants don't make a wave, but their placement is the signal: when reinforcement-learning methods start classifying under H02J 3/17 and H02J 3/004 rather than purely under G06N, it means the patent system is treating AI as grid infrastructure — a control layer that sits on the power network, claimed alongside converters and demand-response arrangements. That is a different posture than "analytics about the grid."
Where the white space likely is: the demand-response and VPP corners are filling, but the same RL-as-controller pattern has obvious unclaimed adjacencies — grid-forming inverter control, microgrid dispatch, transmission-level congestion management. This desk will be watching whether the H02J learning-agent cluster spreads from the edge (homes, EV fleets) toward the core (substations, transmission). The honest qualifier: this is a directional read off two representative grants, not a saturation count — but the cross-class placement is exactly the kind of early signal a landscape report exists to catch.