Energy Forecasting

Energy forecasting predicts load, consumption, generation, or price over horizons from minutes to years. Inputs include historical load, weather forecasts, calendar effects, tariffs, outages, distributed generation, and known industrial schedules. The target is decision-specific: hour-ahead balancing needs low-latency point forecasts, day-ahead operations need peak-aware load forecasts, and planning needs uncertainty bands around electrification or solar adoption.

Framing

This page is the applied version of energy consumption forecasting. Weather covariates and calendar encoding are usually as important as model choice, so feature engineering for forecasting should include temperature lags, heating and cooling degree terms, daylight, holidays, and local business schedules. Evaluation should be by horizon, feeder or region, season, and peak period. For reserve planning, probabilistic forecasting and forecast calibration are more useful than a single MAE.

The UCI household power dataset is a concrete public artifact: it contains 2,075,259 one-minute measurements from a household near Paris between December 2006 and November 2010, with about 1.25% missing rows.

Executed Artifact

To show how much of the load signal lives in weather and calendar features, the example below pits a same-hour-yesterday baseline against a gradient-boosted model that also sees temperature and calendar terms, and it reports the peak-hour error separately because average accuracy can hide missed ramps.

import os
import numpy as np
from sklearn.ensemble import HistGradientBoostingRegressor
 
os.environ["LOKY_MAX_CPU_COUNT"] = "4"
rng = np.random.default_rng(18)
h = np.arange(24 * 21)
temp = 15 + 8 * np.sin(2 * np.pi * (h % 24 - 14) / 24) + 3 * np.sin(2 * np.pi * h / (24 * 7))
load = (
    100
    + 2.8 * np.maximum(temp - 18, 0)
    + 1.7 * np.maximum(10 - temp, 0)
    + 9 * ((h % 24 >= 17) & (h % 24 <= 21))
    + rng.normal(0, 1.2, len(h))
)
train = h < 24 * 18
test = ~train
X = np.c_[h % 24, h % 168, temp]
model = HistGradientBoostingRegressor(random_state=18, max_iter=80).fit(X[train], load[train])
pred = model.predict(X[test])
yesterday = load[h[test] - 24]
mae = lambda y, p: np.abs(y - p).mean()
 
print("yesterday_same_hour_mae_mw", round(mae(load[test], yesterday), 2))
print("weather_calendar_hgb_mae_mw", round(mae(load[test], pred), 2))
print("peak_hour_abs_error_mw", round(abs(load[test][np.argmax(load[test])] - pred[np.argmax(load[test])]), 2))

Observed output:

yesterday_same_hour_mae_mw 2.62
weather_calendar_hgb_mae_mw 1.06
peak_hour_abs_error_mw 0.09

The weather-calendar model cut MAE by more than half and kept the peak-hour miss to 0.09 MW. In grid use, the peak slice would be reported separately because a model can look good on average while underestimating rare ramps.

Failure Modes

Weather forecast error propagates directly into load forecasts. Demand response, new tariffs, electrification, rooftop solar, and economic changes create concept drift. Production forecast monitoring should track missing telemetry, calibration by quantile, ramp errors, and peak-period residuals.

References

• UCI Machine Learning Repository: Individual Household Electric Power Consumption
• Gasparin, Lukovic, and Alippi, Deep Learning for Time Series Forecasting: The Electric Load Case

Nav

Section — Domain Applications

← Demand Prediction in Logistics Predictive Maintenance →