New load forecast code

[springfall2008]

Summary

Adds a new weighted-bucket historical load forecast (enabled with days_previous_auto), and fixes how car‑charging hold is applied in the legacy days_previous averaging.

New: weighted-bucket load forecast (days_previous_auto)

Set days_previous_auto: True in apps.yaml to switch house‑load prediction from the fixed days_previous averaging to a weighted‑bucket forecast. This is more robust when there are gaps in the load history or when usage patterns change (e.g. returning from holiday), because it no longer depends on a small number of specific days all being present and representative.

• Uses all available history within the search window, with no padding when fewer days exist. The window is taken from max(days_previous) (or 7 days if days_previous is not set), capped at 30 days.
• For each forward 5‑minute slot, the historical samples at the same time‑of‑day are combined into a weighted average. Per‑sample weight = weekday × holiday × age:
  • Weekday: 1.0 same weekday as the slot; 0.7 if different but both weekday or both weekend; 0.5 cross (weekday vs weekend).
  • Holiday: ×0.5 when the holiday‑mode state at the moment the sample was recorded does not match today's. Historical holiday state is reconstructed from the recorded holiday_days_left history (via minute_data) and matched per 5‑minute bucket so a mid‑day change of holiday mode is handled correctly.
  • Age: 0.9 for yesterday, reducing by 0.03/day down to a floor of 0.1.
• Zero (missing‑data) buckets are ignored entirely so gaps in the history do not drag the estimate down.
• Produces a cumulative‑from‑midnight forecast that is consumed exactly like the ML load forecast (via load_forecast + load_forecast_only). load_ml takes precedence if both are enabled.

Fix: car‑charging hold applied per day before averaging

The legacy days_previous path averaged the selected days first and only then applied the car‑charging‑hold threshold. A single day's EV charge therefore got diluted below the threshold by the other (lower‑load) days and leaked into the house‑load forecast (e.g. an evening forecast spiking to ~3 kWh when no individual recent day was that high).

get_filtered_load_minute now applies the car/iBoost subtraction and the car‑charging‑hold threshold to each previous day individually and only then weighted‑averages them (new get_filtered_load_window helper). This matches the per‑sample behaviour of the new weighted‑bucket forecast.

Other changes

• The weighted forecast is genuinely cumulative‑from‑midnight (so consumers such as load_today_comparison that read minutes before minutes_now get the right values) and correctly handles midnight crossings — each slot samples whole, distinct days at the slot's time‑of‑day.
• Test harness: the interactive single debug run (--debug) and the debug_cases regression suite now exercise the same recompute path (new redo flag / --noredo), and the plan comparison is value‑based rather than a raw JSON string compare (robust to numpy‑vs‑python float repr).

Tests & docs

• New test_load_forecast_history.py: weekday/age weighting, per‑bucket holiday weighting, zero‑bucket exclusion, short/gappy history, days_previous_auto activation, per‑day car‑charging hold, and cumulative‑from‑midnight / midnight crossing.
• New test_filtered_load_minute.py: get_filtered_load_window / get_filtered_load_minute (raw, car/iBoost subtraction, hold threshold, historical weighted averaging, base load).
• docs/apps-yaml.md: documents days_previous_auto.

[Copilot]

Pull request overview

This PR introduces a new days_previous_auto mode that replaces the fixed days_previous averaging with a weighted “bucketed” historical load forecast built from all available history (up to a capped window), and updates docs/tests accordingly.

Changes:

• Added days_previous_auto config option plus a new weighted historical load forecast implementation (compute_load_forecast_history) and wiring to enable it during sensor fetch.
• Refactored load filtering to apply car/iBoost filtering per-day before weighted averaging, and added new unit tests for the new forecast mode and filtering logic.
• Updated documentation and repo guidance (AGENTS.md), and expanded cspell dictionary.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
docs/apps-yaml.md	Documents the new days_previous_auto weighted historical load forecast behavior and weighting factors.
apps/predbat/unit_test.py	Registers new test modules in the test runner.
apps/predbat/tests/test_single_debug.py	Makes expected-vs-actual plan comparison value-based instead of raw JSON string comparison.
apps/predbat/tests/test_load_forecast_history.py	Adds coverage for weighted historical forecast logic and holiday reconstruction behavior.
apps/predbat/tests/test_filtered_load_minute.py	Adds coverage for per-window/per-day filtering and averaging behavior.
apps/predbat/predbat.py	Initializes the new load_forecast_history flag in reset.
apps/predbat/fetch.py	Implements the weighted-bucket historical forecast, holiday history reconstruction, and refactors load filtering.
apps/predbat/const.py	Adds LOAD_FORECAST_HISTORY_MAX_DAYS constant (30-day cap).
apps/predbat/config.py	Adds schema entry for days_previous_auto.
AGENTS.md	Adds repository guidance for agents (tests, pre-commit, architecture notes).
.cspell/custom-dictionary-workspace.txt	Adds “backtest” and “holdout” to spell-check dictionary.

[Copilot]

Pull request overview

Copilot reviewed 11 out of 12 changed files in this pull request and generated 3 comments.