Load ML changes to allow up to 28 days of data for training

[springfall2008]

#3388

[Copilot]

Pull request overview

This PR enhances the ML Load Prediction component to support training on up to 28 days of historical data (previously limited to 7 days). The key improvement is that fine-tuning now uses the full dataset with time-weighted sampling rather than just the last 24 hours, which prevents catastrophic forgetting while still prioritizing recent patterns through exponential decay weighting.

Changes:

• Increases default historical data fetch from 7 to 28 days with new configurable parameter load_ml_max_days_history
• Removes the distinction between initial training and fine-tuning data ranges - both now use full available dataset
• Updates Temperature API to fetch 28 days of historical temperature data instead of 7
• Adds comprehensive documentation explaining the time-weighted sampling approach and rationale for using full dataset during fine-tuning

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
docs/load-ml.md	Documents new 28-day configurable limit, explains time-weighted sampling (7-day decay), and clarifies that fine-tuning uses full dataset
apps/predbat/load_predictor.py	Removes conditional logic for fine-tuning vs initial training data ranges; both now use full available dataset with consistent time-weighted sampling
apps/predbat/load_ml_component.py	Adds load_ml_max_days_history parameter (default: 28), updates data fetching to use configurable days instead of hardcoded 7
apps/predbat/components.py	Registers new config parameter and updates Temperature API URL to fetch 28 days of past data

[Copilot]

The use of max() with ml_max_days_history is confusing. The parameter name suggests it's an upper limit, but max(ml_max_days_history, ml_min_days) treats it as the target value to fetch. Consider renaming to ml_days_history or ml_target_days_history to better reflect its purpose as the target number of days to fetch (which also serves as the maximum). Alternatively, the logic could be more explicitly written as min(max(ml_min_days, available_days), ml_max_days_history) if it's truly meant to be a cap.

Suggested change

	days_to_fetch = max(self.ml_max_days_history, self.ml_min_days)
	days_to_fetch = min(max(self.ml_min_days, 1), self.ml_max_days_history)