nesoai/portugal-electricity-demand-forecaster

Electricity Demand Forecaster for Portugal (PT)

High-accuracy calibrated quantile LightGBM model for forecasting Portugal demand. Resolution: 15-minute intervals. Trained on multi-year data (2023–2026) from ENTSO-E, featuring multi-scale lags, cyclical encodings, and conformal calibration for well-calibrated 80% prediction intervals ($P10, P50, P90$).

Model Highlights

• Country / Zone: Portugal (PT)
• Target: Electricity Demand in MW
• Resolution: 15-minute intervals
• Outputs: Point forecast ($P50$), 80% prediction interval ($P10$ to $P90$)
• Algorithm: LightGBM Multi-Quantile Regressor with Conformal Calibration & Monotonicity
• Dataset: ENTSO-E European Transparency Platform (Zone: PT)
• Training Samples: 96,665 observations (2023–2026)

Performance & Benchmark Comparison

Evaluated on out-of-sample test sets against official seasonal persistence benchmarks:

Metric	LightGBM Forecaster	7-Day Seasonal Persistence	Improvement
MAE	45.528 MW	716.312 MW	+93.6%
RMSE	90.863 MW	—	—

• WAPE: 0.72% | P10 Pinball Loss | 33.950 | — | — | | P90 Pinball Loss | 14.355 | — | — | | P10–P90 Interval Coverage | 78.4% | — | Target: 75–85% | | Winkler Score | 483.052 | — | — |

Quickstart: Python Inference

import pandas as pd
from huggingface_hub import hf_hub_download
import joblib

# 1. Download model artifacts
model_path = hf_hub_download(repo_id="ORGANIZATION/portugal-demand-forecaster", filename="models.joblib")
models = joblib.load(model_path)

# 2. Predict P10, P50 (point), and P90 quantiles
X_test = pd.read_csv("sample_input.csv")
p10 = models[0.1].predict(X_test)
p50 = models[0.5].predict(X_test)
p90 = models[0.9].predict(X_test)

print("Forecast Point Estimate:", p50[:5])
print("80% Lower Bound (P10):", p10[:5])
print("80% Upper Bound (P90):", p90[:5])

Features Used

The model uses 37 leakage-safe features:

• hour
• quarter
• day_of_week
• day_of_year
• month
• is_weekend
• is_holiday
• is_morning_peak
• is_evening_peak
• sin_hour
• cos_hour
• sin_day_of_week
• cos_day_of_week
• sin_day_of_year
• cos_day_of_year
• lag_1h
• lag_2h
• lag_3h
• lag_4h
• lag_24h
• lag_48h
• lag_7d
• lag_14d
• diff_1h
• diff_2h
• diff_24h
• diff_7d
• acceleration_1h
• ema_4step
• ema_12step
• rolling_std_4step
• rolling_mean_24h
• rolling_std_24h
• rolling_min_24h
• rolling_max_24h
• rolling_mean_7d
• rolling_std_7d

Intended Use & Advisory

This model is intended for research, energy market analytics, grid load planning, and educational forecasting demonstrations. It is advisory only and not intended for automated trading execution or real-time grid dispatch.

Citation & Attribution

Data published under the ENTSO-E Transparency framework:

• Transparency Platform: https://transparency.entsoe.eu/