DailyProfileTransformer

Copies each local day’s values of a column onto every row of that day. With hourly data, price becomes price_h00 … price_h23: on every row of day D, price_h05 is the price at 05:00 on D. With 15-minute data the columns are price_s00 … price_s95.

The library fits one model per hour, and each row only carries its own hour’s values. Lagging profile columns by days makes cross-hour inputs possible, such as LEAR’s “all 24 prices of the previous day” (see Predictor Presets).

Basic Usage

from epftoolbox2.data.transformers import DailyProfileTransformer, LagTransformer

df = DailyProfileTransformer(columns=["price"]).transform(df)
df = LagTransformer(columns=[f"price_h{h:02d}" for h in range(24)], lags=[1, 2, 6], freq="day").transform(df)

Parameters

Parameter	Type	Default	Description
columns	str | List[str]	required	Column name(s) to spread over the day
freq	str	"1h"	Slot length of the data: "1h" adds 24 columns per input column, "15min" adds 96

epftoolbox2.presets.profile_columns("price", freq="1h") returns the column names.

Information Set

A row of day D only gets values of day D. Whether they are known when a forecast is issued on D depends on the column:

• Day-ahead prices and day-ahead forecasts for D are published on D-1, so price_h05 on the rows of D is known.
• Realised values, such as load_actual, are not known for the rest of D in the morning of D. Use their profiles only lagged: load_actual_h05_d-1.

Caution

Leading a profile (price_h05_d+1) looks up the prices being forecast. Only lead columns that are known in advance.

DST Days

On a time-zone-aware index the day of the spring switch has no 02:00, so price_h02 is NaN on its rows, and on the autumn switch day price_h02 takes the first 02:00. A warning suggests running DSTNormalizeTransformer first, which gives every day 24 slots:

df = DSTNormalizeTransformer(freq="1h").transform(df)
df = DailyProfileTransformer(columns=["price"]).transform(df)

A NaN profile column makes the model skip every forecast whose row uses it, so without normalisation a model with price_h02_d-1 loses all forecasts issued on the day after the spring switch.

On a UTC index the “days” are UTC days, and the transformer warns.