ResampleTransformer

Brings every column of a frame to one regular frequency. It aggregates when a column is finer than the target (15 min → 1 h) and repeats a value over its own period when the column is coarser (1 h → 15 min). It does not fill missing data and never uses a later value to fill an earlier slot.

Basic Usage

from epftoolbox2.data.transformers import ResampleTransformer

transformer = ResampleTransformer(freq="1h")
df = transformer.transform(df)

Parameters

Parameter	Type	Default	Description
freq	str	"1h"	Target pandas frequency ("15min", "1h", "1D", …)
agg	str | dict	"mean"	Aggregation used when downsampling: mean, sum, first, last, min, max, median, or a dict {column: aggregation}
fill	str | None | dict	"ffill"	How slots created by upsampling are filled: "ffill" (repeat the value inside its own period) or None (leave them empty), or a dict {column: fill}

In a dict, columns that are not listed use the "default" entry, or "mean" / "ffill" when there is none:

# Prices and load are averaged, energy is summed
ResampleTransformer(freq="1h", agg={"default": "mean", "energy_mwh": "sum"})

# Keep the hourly price only at the full hour when going to 15 minutes
ResampleTransformer(freq="15min", fill={"price": None})

Non-numeric columns (e.g. weekday_name) use "first" when the aggregation is numeric-only (mean, sum, median).

How it works

Every value is the value of a period that starts at its timestamp and lasts as long as the column’s own spacing: 15 minutes, one hour, or one local calendar day for daily columns. The spacing is worked out per column from the timestamps of its non-null values, so a frame that mixes hourly prices, 15-minute load and daily A31 values is handled column by column. A column can even change resolution part-way (hourly day-ahead prices before the 15-minute switch, 15-minute prices after it).

Upsampling (column coarser than freq): each finer slot inside a value’s period gets that value, so an hourly price becomes a step: 10:00, 10:15, 10:30 and 10:45 all hold the 10:00 price. Dummies stay 0/1. The last period is completed, so an hourly day becomes 96 quarter-hours. For sum columns the value is split evenly across the slots, so totals are kept. With fill=None the new slots stay empty.

Downsampling (column finer than freq): the values inside each target period are combined with agg. A period that is not completely covered gives NaN: a missing quarter-hour makes its hour NaN, and the partial first and last local days make a daily mean NaN. You never get a partial sum or a sum of zero for an empty period.

Nothing else is filled. A value missing in the source (an outage, a price that isn’t published yet) stays missing, and nothing is extrapolated past a column’s last value. Gaps therefore stay visible to validators and are dropped by the models instead of being scored against invented actuals.

Pipeline order: Timezone → Resample → Lag

from epftoolbox2.pipelines import DataPipeline
from epftoolbox2.data.sources import EntsoeSource
from epftoolbox2.data.transformers import TimezoneTransformer, ResampleTransformer, LagTransformer

pipeline = (
    DataPipeline()
    .add_source(EntsoeSource(country_code="PL", api_key="...", type=["load", "price"]))
    .add_transformer(TimezoneTransformer(target_tz="Europe/Warsaw"))  # local time, daily values on local midnight
    .add_transformer(ResampleTransformer(freq="1h"))                  # one regular grid
    .add_transformer(LagTransformer(columns=["price"], lags=[1, 7], freq="day"))
)

1. TimezoneTransformer first, so that daily bins (freq="1D") and daily columns follow local calendar days and daily values sit on local midnight. Resampling a UTC frame to "1D" aggregates UTC days, and resampling daily columns before the timezone step makes TimezoneTransformer raise.
2. ResampleTransformer next, to put every column on the same grid.
3. LagTransformer last, on the regular local grid.

What happens on DST days

After TimezoneTransformer, a local day has 23 hours on the spring switch day and 25 on the autumn one. ResampleTransformer keeps that: the spring day has 23 hourly rows (92 quarter-hours) and no 02:00 row, because that local time doesn’t exist. The autumn day has 25 hourly rows (100 quarter-hours) with 02:00 twice. Nothing is filled in for the “missing” hour, and a daily value covers all 23 or 25 hours of its day.

Notes

• The input must have a DatetimeIndex without duplicate timestamps. An unsorted index is sorted.
• Timestamps are period starts (ENTSO-E convention).
• Output values are not rounded.
• Numeric columns come out as float64.