feat(dynamic-load): cancel a car's slots and IOG cheap rate while it is not charging, with an option to distrust Intelligent slots until seen charging

[springfall2008]

An alternative to #5110, and now also covers its main case (not trusting future dispatches) through an opt-in switch.

Problem

Octopus Intelligent dispatch slots are treated as cheap for the house battery whether or not the car actually charges in them. If the car doesn't draw, Octopus may bill that import at the full rate (#5065), and Octopus can also withdraw a slot before it happens.

#5110 tackles this by distrusting future dispatches through the rates only. That leaves car_charging_slots untouched, so "Hold for car" still holds the battery during a slot the plan no longer trusts. At its strictest setting, with car_charging_from_battery on, the battery drains into the car during a genuine dispatch.

Dynamic load already cancelled a car's future slots when the load showed the car wasn't charging. But it ran after the rates were built, so the dispatch kept its cheap rate.

1. Dynamic load: cancel slots, and their cheap rate, while the car isn't charging

This part is gated by the existing switch.predbat_metric_dynamic_load_adjust. It only applies to slots built by Octopus Intelligent charging. Predbat-led charging is never cancelled: there, Predbat itself decides when the car charges, and its car_charging_slot sensor drives the charger, so cancelling a slot would stop the very charge the check is waiting to see.

The car check moves into fetch_sensor_data(): after the Octopus slots are built, before the rates. It decides every car in one call. While an Intelligent car is inside one of its slots and not charging:

• Car slots: every slot ending after now is set to 0 kWh. This releases "Hold for car" and removes the predicted car load.
• Predbat's own discount: rate_add_io_slots() skips that car's dispatch minutes from now onwards.
• Feed discount: a discount the rate feed already delivered (io_adjusted) is set back to rate_max_base before replication, rounded out to whole 30-minute rate periods. Three kinds of minute are left alone:
  • elapsed minutes, so today's cost stays true;
  • the fixed 23:30-05:30 window, which is cheap by tariff, not by dispatch;
  • minutes where another car that is still trusted has a dispatch.

Evidence that the car isn't charging:

Source	Used when	Must hold for
car_charging_now	it's a real entity. A static literal, an entity HA doesn't have, and unknown/unavailable are all no evidence either way	2 minutes
Low load	there's no sensor and the car is inside the CT clamp (car_energy_reported_load)	10 minutes
—	otherwise	no detection

Reset: the decision is worked out again every cycle. It ends as soon as the car leaves its slots or starts charging again.

Poll: update_time_loop (every 15 seconds) asks for a replan as soon as the state is due to change, instead of waiting up to 5 minutes for the next cycle. It makes the same decision as the check, on the same floored, clock-skewed minute, so the two can't disagree at a slot boundary.

dynamic_load() keeps the high-load baseline unchanged. The load classification is now a shared helper.

2. switch.predbat_octopus_intelligent_trust_slots (expert mode, default On)

• On: everything in section 1. A slot is trusted until the car is seen not charging in it.
• Off: reversed. A car's Octopus Intelligent slots start untrusted: 0 kWh, and no cheap daytime rate from now onwards.
  • As soon as the car is seen charging inside a slot, that slot and every later one are trusted straight away.
  • They're cancelled again if the car stops charging for the usual wait, or when the slot ends. So each separate dispatch has to be confirmed.
  • The overnight rate stays cheap, as it's part of the tariff.
  • Off works whether or not dynamic load is on. It never affects slots Predbat planned itself.
  • It only acts on the car plan that Octopus Intelligent charging builds. With octopus_intelligent_charging Off, the dispatch rates are used as before and a warning is logged.
  • If a car has no sensor and is outside the CT clamp, nothing can ever confirm it charging. Predbat logs a warning whenever the set of such cars changes.

Both directions come from one function, dynamic_load_car_target():

Where now is	What the car shows	Result
Outside any slot	anything	the default: trusted if On, untrusted if Off
Inside a slot	charging	trusted, immediately
Inside a slot	not charging, for the wait	cancelled
Inside a slot	no evidence	unchanged

3. Cancelled slots stay visible, with a "?"

A cancelled slot keeps the kWh it had in a new kwh_cancelled field on the slot itself. That means it travels with the existing deep copies in compare.py and the yesterday plan, and disappears when the slots are rebuilt each cycle.

• Published car plan: a cancelled slot has 0 kWh and 0 cost, plus its kwh_cancelled, in the planned attribute of binary_sensor.predbat_car_charging_slot.
• HTML plan: the car column shows 3.5? on pale yellow. If another car is charging in the same step, it shows both, e.g. 2.0 +3.5?.
• JSON: a new car_charging_cancelled field, which the web page shows the same way.

The planned car_charging value and the totals are unchanged.

Behaviour change for existing dynamic load users

• The load-based cancellation now needs 10 minutes of low load. Before, it was two consecutive checks. It still needs the car inside the CT clamp, as before.
• For Intelligent slots it also removes the dispatch's cheap rate, not just the car load.
• A car with a real car_charging_now sensor is judged on that sensor, not on the load.
• Slots Predbat plans itself (Predbat-led charging) are no longer cancelled on low load. The old dynamic_load() zeroed them too.

Review round 1

• Fixed:
  • a missing car_charging_now entity read as "not charging";
  • the feed-discount strip used exact dispatch boundaries instead of whole rate periods;
  • compare.py runs applied a cancellation that was out of date;
  • a missing bounds check.
• Discussed in the threads: rate_max_base as the restored price; the CT clamp condition; the poll's replan cost.

Review round 2

• Fixed:
  • the poll now reads a sensor entity cached once per cycle, so a single sensor shared by two cars no longer logs a set-up warning every 15 seconds;
  • a cancelled car's "?" now shows alongside another car's live charging;
  • the trust switch's precondition (Octopus Intelligent charging On) is now documented, with a warning;
  • the dead load_last_car_slot state has been removed.
• Predbat-led charging: an earlier version of this PR also cancelled Predbat-planned slots, in a second check that ran before publish_car_plan(). That could switch off the slot sensor driving the charger, so the charge the check was waiting for never started. Cancelling is now limited to Octopus Intelligent slots.
• Checked: the poll across midnight. It reads the previous cycle's slots and its midnight_utc together, so they stay on the same axis. Test 18b pins this.

Review round 3

• Fixed:
  • in compare.py runs, the rates now follow the same decision as the slots (dynamic_load_car_effective);
  • no load history no longer counts as low load;
  • an unknown reading restarts the grace timer;
  • a cancelled slot's cost is zeroed, and kwh_cancelled is published in the car plan's planned attribute;
  • the "never trusted" warning is limited to Octopus Intelligent cars;
  • the feed strip logs only when what it strips changes.
• Discussed in the threads: deriving the night window from the tariff; rate-limiting poll-triggered replans.

Tests

test_dynamic_load_car_not_charging has 41 scenarios. They cover:

• grace timing for each evidence source, and which source takes priority
• unknown readings, missing entities and static literals
• the CT clamp condition and the midnight load reset
• both reset paths (leaving the slot, charging again)
• compare.py's save=False runs
• Predbat-planned cars never being cancelled, and a cancellation being cleared when a car's slots disappear
• rate withholding
• feed-discount stripping: past minutes, 30-minute rounding, another car's dispatch, and a dispatch straddling 23:30
• the poll, including the slot-boundary case, across midnight, and with no per-poll warnings
• the trust switch: its default, confirmation, re-cancel, the slot ending, own slots untouched, no evidence, the poll, and the warning
• kwh_cancelled

test_plan_json_rate_adjust checks the 3.0? HTML cell, the JSON field, the unchanged totals, and the two-car 2.0 +3.0? cell.

To check the tests themselves, deliberately broken versions of the logic were confirmed to make the suite fail (over 30 mutations across the rounds). test_dynamic_load_car_slot_cancellation is updated, because dynamic_load() no longer cancels slots itself.

./run_all --quick: all passed. pre-commit: clean.

🤖 Generated with Claude Code

[chalfontchubby]

This isn't strictly an alternative to 5110 in that it does not close the hole of planning to use slots that go away. I do want to land that aspect of future trust.
At a first glance this does have merit once you are in a slot, but it only closes one of the concerns we have - perhaps the one you indicate is a lower risk based on reconciling bills.

[chalfontchubby]

This does seem a better solution for the subset of the problem it addresses but it doesn't close the issue, and won't shut me up :-)

[springfall2008]

@chalfontchubby I guess an add on would be not to trust future IOG slots, so don't mark them as low rate on the plan until the car is seen charging, this is just flipping the default of this PR?

[springfall2008]

Automated review failed to complete for this PR - see the triage bot's logs for details. The run exited cleanly but posted nothing, so the review step itself did not complete. Not retrying automatically; remove BOT_FAILED and re-add BOT_REVIEW to try again.