Skip to contents

Fit a Kernel Wavelet Functional model for daily load curves

Source: R/elcf4r_fit_kwf.R
elcf4r_fit_kwf.Rd

Fit a day-ahead Kernel Wavelet Functional (KWF) model on ordered daily load curves. The implementation computes wavelet-detail distances on the historical context days, applies Gaussian kernel weights, restricts those weights to matching calendar groups when available, and can apply the approximation/detail correction used for mean-level non-stationarity.

Usage

elcf4r_fit_kwf(
  segments,
  covariates = NULL,
  target_covariates = NULL,
  use_temperature = FALSE,
  wavelet = "la12",
  bandwidth = NULL,
  use_mean_correction = TRUE,
  group_col = NULL,
  holidays = NULL,
  weights = NULL,
  recency_decay = NULL,
  temperature_bandwidth = NULL
)

Arguments

segments

Matrix or data frame of past daily load curves (rows are days, columns are within-day time points) in chronological order.

covariates

Optional data frame with one row per training segment. When present, the function looks for deterministic grouping information in context_group, kwf_group, calendar_group, or the column named by group_col. If no explicit group column is present, groups are derived from date and holidays, or from dow as a fallback.

target_covariates

Optional one-row data frame describing the day to forecast. When it contains date, the previous day is used as the context day for calendar grouping, which matches the residential KWF protocol for pre-holiday handling.

use_temperature

Deprecated and ignored. Kept for backward compatibility with earlier package examples.

wavelet

Wavelet filter name passed to wavelets::dwt(). Defaults to "la12", the least-asymmetric filter. If the series is too short for the requested filter, the function falls back to "haar".

bandwidth

Optional positive bandwidth for the Gaussian kernel on wavelet distances. If NULL, it is inferred from the distances to the last observed segment.

use_mean_correction

Logical; if TRUE, apply the approximation/detail correction used for mean-level non-stationarity.

group_col

Optional column name containing precomputed KWF groups in covariates and target_covariates.

holidays

Optional vector of holiday dates used by elcf4r_calendar_groups() when deterministic groups are derived from date.

weights

Optional numeric prior weights of length nrow(segments). Only the first nrow(segments) - 1 values are used in the historical pairing step.

recency_decay

Optional non-negative recency coefficient applied as an exponential prior on the historical context days.

temperature_bandwidth

Deprecated and ignored. Kept only for backward compatibility with older examples.

Value

An object of class elcf4r_model with method = "kwf".

Examples

id1 <- subset(
  elcf4r_iflex_example,
  entity_id == unique(elcf4r_iflex_example$entity_id)[1]
)
daily <- elcf4r_build_daily_segments(id1, carry_cols = "participation_phase")
fit <- elcf4r_fit_kwf(
  segments = daily$segments[1:10, ],
  covariates = daily$covariates[1:10, ],
  target_covariates = daily$covariates[11, , drop = FALSE]
)
length(predict(fit))
#> [1] 24