Electricity Load Forecasting for Ireland

Electricity demand forecasting requires models that perform reliably across multiple forecast horizons. Short-term forecasts capture immediate load behavior, while day-ahead and week-ahead horizons require stronger temporal structure, seasonality handling, and covariate evaluation.

The project focused on testing whether weather-augmented machine learning improves forecasting performance beyond seasonal and autoregressive baselines.

Electricity demand was sourced from ENTSO-E at hourly resolution. Weather covariates were collected from NASA POWER and aligned to the demand time index.

• Hourly electricity demand and weather series aligned on a shared timeline
• Lag and calendar features derived only from information available at forecast origin
• Weather variables joined with careful temporal ordering to avoid future leakage
• Feature engineering preserved train/validation chronology for each horizon

Exact record counts depend on the published ENTSO-E and NASA POWER extracts used in the repository workflow.

1. 01

   Data ingestion and alignment

   Collected and aligned electricity demand with weather variables at hourly resolution while preserving temporal ordering.

2. 02

   Feature engineering

   Prepared lagged demand features, calendar variables, seasonal indicators, and weather covariates without introducing future leakage.

3. 03

   Model benchmarking

   Compared statistical, machine learning, and deep learning approaches under a consistent evaluation protocol.

4. 04

   Rolling-origin validation

   Evaluated models across t+1, t+24, and t+168 horizons using walk-forward splits that respect the time axis.

5. 05

   Performance comparison

   Compared model behavior across horizons to understand where weather variables add measurable value.

Each model was included to answer a specific benchmarking question—from whether complexity beats seasonal repetition to whether weather covariates improve multi-horizon forecasts.

The project used rolling-origin / walk-forward validation to preserve temporal integrity. Instead of randomly splitting data, each evaluation step trained on historical observations and tested on future periods. This avoids information leakage and better represents real forecasting conditions.

• Multi-horizon forecasting pipeline
• Weather-augmented feature set
• Benchmark comparison across statistical, ML, and deep learning models
• Leakage-aware validation workflow
• Reproducible codebase

• Used rolling-origin evaluation instead of random split
• Benchmarked against seasonal naive to avoid weak baseline comparison
• Evaluated multiple horizons instead of only one-step forecasting
• Used weather covariates to test external signal contribution
• Preserved reproducibility through structured scripts and notebooks

This project demonstrates time-series validation awareness, multi-horizon forecasting design, model benchmarking, feature engineering, and the ability to evaluate whether external covariates meaningfully improve forecasting performance.

• Results depend on data quality and weather–demand alignment
• National/system-level forecasts may hide local demand variation
• Additional external variables could improve future models
• Production deployment would require monitoring and retraining