case study 05

Electricity Load Forecasting (Time Series)

forecasts a full day of 15-minute load (96 steps), and rebuilds a research algorithm as an R package.

scikit-learn
XGBoost
SARIMA
LSTM
R

time-series-electricity-forecasting / WNN_Package/R/wnn_forecast.R

dists <- apply(model$X, 1, function(row) euclidean_distance(row, window))
idx <- order(dists)[1:model$k]
weights <- 1/(dists[idx] + 1e-8)
weights <- weights / sum(weights)
sum(weights * model$y[idx])

why: most demos stop at one-step-ahead forecasts, which hide how error compounds. i wanted a model that still held up 96 steps out.
how it works: exponential smoothing, SARIMA, random forest, XGBoost, SVR, MLP and LSTM are compared with TimeSeriesSplit on lag, rolling and calendar features. the weighted nearest neighbours method (Talavera-Llames et al., 2016) is implemented from the paper as an installable R package.
what i learned: recursive multi-step forecasting punishes small early errors, feature engineering mattered more than model choice.
results: a final 96-step forecast, plus the wnnforecast R package with roxygen docs and a vignette.

still testing how far the recursive horizon can stretch before error compounds too much.