normalization.thermosensitive — explanation

Goal

Compare an offer using consumption recorded under abnormal climate (e.g. a

cold winter) against a "normal" climate baseline. The method removes the

weather-driven variance from the consumption series so the offer comparison

isolates the underlying tariff structure, not the weather.

A piecewise-linear heating-degree-day model on daily aggregates:

```

daily_kwh ≈ base_kwh + α · HDD(T_day)

HDD(T) = max(0, base_temp − T)

```

base_kwh — the non-thermosensitive baseload (lighting, appliances, …)

α — the thermosensitive slope (kWh per °C-day below base_temp)

base_temp — the heating threshold (default 17 °C, fittable)

If (α, base_temp) are supplied, skip the fit.

Else fit by ordinary least squares against (T_day, daily_kwh):

- For each candidate base_temp ∈ {14, 15, 16, 17, 18, 19, 20}, regress

daily_kwh on HDD(T_day; base_temp). Pick the base_temp with the

highest R²; report its (α, base_kwh).

Normalize — for each day, replace observed T_day with the

corresponding T_normal_day and recompute:

```

daily_kwh_normal = daily_kwh − α · (HDD(T_day) − HDD(T_normal_day))

```

The annual total is then sum(daily_kwh_normal).

Linear model — does not capture cooling load (no CDD term).

A 2-year span is the minimum recommended input for a stable fit; with one

year the fit is noisy.

base_temp search grid is coarse on purpose: finer grids overfit on small

samples.