fix(car): low-load test needs live load power to be low too before cancelling an Intelligent dispatch

[springfall2008]

Fixes #5461.

Problem

With octopus_intelligent_dynamic on and no car_charging_now sensor, the slot confirmation falls back to the house load. It cancelled a genuine dispatch while a 7.4 kW car was charging in view of the CT: the 5-minute load figure read 4.80 kW against a 5.4 kW cut-off, at a moment when the inverter reported 7410 W. Once the slots were cancelled the battery hold was released and the battery discharged into the car.

The 5-minute figure lags a car that is still ramping up, and a cloud inverter whose readings arrive late. The load path has no start band, so the ramp window itself can start the grace clock and one more marginal window cancels.

Change

• Live load must be low too. In dynamic_load_car_evidence(), a low 5-minute figure only counts as "not charging" when the live load_power total is also under the cut-off.
• Mixed evidence is no evidence. A low 5-minute figure with live load that is not low, or has no reading, returns "no evidence": it neither cancels the car nor resumes it, and it resets the grace clock.
• Cut-off lowered from 90% to 85% of car_charging_threshold (5.1 kW on the 6 kW default), as DYNAMIC_LOAD_CAR_LOW_FACTOR. This gives margin for a load sensor that under-reads while the battery is grid-charging.
• Installs without load_power keep the 5-minute figure alone, as before.
• Logging. Each low window logs the 5-minute figure, the cut-off and the live figure.

The live figure is read from the load_power sensors directly and summed across inverters, because dynamic_load_car_check() runs before fetch_inverter_data() and the inverter objects still hold the previous cycle's value. If any inverter's sensor is unknown, unavailable or missing, there is no total.

Things to review

• dynamic_load_classify() is shared with dynamic_load(), so the 85% factor also moves the published load_energy_last_period status: a load between 85% and 90% of the threshold now reads "baseline" where it read "low". The battery-rate 0.9 in the same test is unchanged.
• A cancelled car resumes on the next 5-minute figure, not on the live reading alone.
• Confirming the half hour (keeping its cheap rate once the car stops) still needs load at the full car_charging_threshold. Not changed here.
• Not covered: a "live" value that is itself minutes stale and was taken during the ramp reads low on both measures. Skipping the first window after the dispatch starts would close that; it is left out of this PR.

Testing

• Tests 60-66 in tests/test_dynamic_load_car.py (./run_all --test dynamic_load_car_not_charging): the 4.80 kW / 7410 W case from the issue, both figures low, two inverters summed, unavailable / partly unavailable / missing sensors, no load_power, a cancelled car with high live load, and the 85% boundary on both figures. They failed before the change.
• The test fixture's apps.yaml sets load_power, so the module removes it for the older scenarios, which judge the 5-minute figure alone.
• ./run_all --quick: all tests passed (4 slow tests skipped).
• pre-commit clean on the changed files.

Docs: docs/car-charging.md, "Checking Intelligent dispatches against the car".

🤖 Generated with Claude Code

[Copilot]

🟡 Changes recommended

Live-load aggregation omits configured load_power_N sensors, allowing false cancellations on split-load installations.

1 open finding

• Include all configured load sensors in live load calculation · New

What changed in this PR

Prevents false Intelligent dispatch cancellations by requiring both averaged and live load to be low.

Changes:

• Adds an 85% low-load threshold and live-load cross-check.
• Treats mixed or unavailable live readings as no evidence.
• Adds documentation and regression tests.

File	Description
apps/​predbat/​const.py	Defines the 85% factor.
apps/​predbat/​plan.py	Implements live-load evidence checks.
apps/​predbat/​tests/​test_dynamic_load.py	Updates threshold commentary.
apps/​predbat/​tests/​test_dynamic_load_car.py	Adds live-load regression scenarios.
docs/​car-charging.md	Documents dispatch-check behavior.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.