Fix ML load inflation: skip clean_incrementing_reverse for power sensors

[amasolov]

Summary

minute_data_load unconditionally applied clean_incrementing_reverse to every sensor it loaded. This function is designed for cumulative kWh counters — it detects "resets" (where the counter drops back to 0) and stitches them into a monotonically increasing series.

When applied to load_power sensors that report instantaneous watts (fluctuating values like 300 W → 900 W → 400 W), clean_incrementing_reverse misinterprets every natural drop as a "reset" and accumulates all positive increments into an ever-growing cumulative series. A realistic ~600 W signal becomes a series peaking at ~20,000+ W.

fill_load_from_power then integrates this vastly inflated "power" data to fill zero-load periods, injecting far too much energy and shifting the entire cumulative load curve upwards. The result: ML actual load showing ~2x real consumption, causing the ML model to train on inflated data and over-predict future load.

Changes

• fetch.py: Add clean_increment parameter to minute_data_load (default True for backward compatibility). Pass clean_increment=False when fetching load_power data in both the standard and GE Cloud paths.
• load_ml_component.py: Same fix for the load_power fetch in _fetch_load_data.
• tests/test_fill_load_from_power.py: Add regression test that demonstrates the inflation mechanism — shows ~2.4x total energy inflation when distorted power data is fed to fill_load_from_power.

Root cause

load_power sensor (instantaneous W):   600, 900, 400, 800, 300, 700, ...
                                           ↓
         clean_incrementing_reverse (bug: treats drops as resets)
                                           ↓
distorted "power" data:                600, 1500, 1500, 2300, 2300, 3000, ...
                                           ↓
         fill_load_from_power (integrates inflated values)
                                           ↓
cumulative load inflated by ~2x

Test plan

• New regression test test_fill_load_from_power_distorted_power_inflation passes — demonstrates 2.39x inflation with distorted data
• All existing test_fill_load_from_power tests pass (no regressions)
• All test_load_ml tests pass (27/27)
• unit_test.py --quick passes (pre-existing history_attribute Test 11 failure unrelated)
• Pre-commit checks pass (ruff, black, cspell, trailing whitespace)

Fixes #3692

Made with Cursor

[Copilot]

Pull request overview

Fixes LoadML “actual load” inflation by preventing cumulative-counter cleanup logic from being applied to instantaneous load_power sensors, avoiding distorted power series that over-inflate fill_load_from_power gap fills and ML training data.

Changes:

• Added a clean_increment parameter to Fetch.minute_data_load (defaulting to True) and threaded it into the underlying minute_data(...) call.
• Updated both core fetch flow and the LoadML component to fetch load_power with clean_increment=False.
• Added a regression test demonstrating the distortion/inflation mechanism when clean_incrementing_reverse is incorrectly applied to fluctuating power readings.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
apps/predbat/fetch.py	Adds clean_increment flag to minute_data_load and disables increment-cleaning when loading load_power for gap filling.
apps/predbat/load_ml_component.py	Fetches load_power history with clean_increment=False before calling fill_load_from_power.
apps/predbat/tests/test_fill_load_from_power.py	Adds regression coverage proving distorted “power” can inflate cumulative load totals and ensures correct behavior with undistorted power.