a-niko22/nbeats-energy-forecaster

N-BEATS Energy Price Forecaster

Generic (univariate) N-BEATS model for short-term Dutch day-ahead electricity price forecasting. Part of the OpenRemote Energy Price Forecasting research project, evaluated as a baseline alongside Prophet and an Encoder-Decoder Transformer (Nazim112/nl-energy-forecaster).

Model details

• Architecture: Generic N-BEATS — 6 stacked residual blocks, each a 4-layer MLP (hidden size 256)
• Input: last 168 hourly price values (EUR/MWh), univariate — no exogenous features
• Output: next 48 hourly price values (EUR/MWh)
• Parameters: ~1.78M
• Framework: PyTorch

Training

• Dataset: CitrusBoy/EnergyPriceForecasting (2015–2025, Dutch day-ahead prices + weather/load features, though this model only uses price)
• Split: 85% train / 15% test, chronological
• Lookback / horizon: 168 / 48 hours
• Window stride: 6 hours (subsampled for faster training)
• Optimizer: Adam, lr=1e-3
• Early stopping: patience=4 on validation loss, stopped at epoch 10/15

Test set results

Metric	Value
MAE	25.06 EUR/MWh
RMSE	37.44 EUR/MWh

Usage

from huggingface_hub import snapshot_download
import sys, numpy as np

repo_dir = snapshot_download("a-niko22/nbeats-energy-forecaster")
sys.path.insert(0, repo_dir)
from predict import load_model, predict

bundle = load_model(repo_dir)

# X: last 168 hourly prices (EUR/MWh), unscaled, most recent value last
X = np.array([...])  # shape (168,)
forecast = predict(bundle, X)  # shape (48,) — next 48 hours, EUR/MWh

Limitations

• Univariate: does not use weather, load, or generation features, unlike the NBEATSx variant used elsewhere in this project.
• Trained on data through 2025; forecasts for periods with unusual market conditions (e.g. extreme negative-price events) may be less reliable.