This project focuses on comparing multiple forecasting models to predict electricity demand, aiming to improve:
- Backup power planning
- Energy storage management
- Grid stability
Accurate forecasts are crucial for optimizing intermittent renewable sources like wind and solar and supporting energy market decisions.
- Source: https://huggingface.co/datasets/EDS-lab/electricity-demand
- Content: Time series data on electricity consumption and weather
- Scope: Commercial and residential buildings in Europe and North America
- Frequency: 15-minute, 30-minute, and 1-hour intervals (Jan 2011 – Dec 2017)
- Parametric Models:
- ARIMA/SARIMA (univariate)
- ARIMAX/SARIMAX (multivariate with exogenous variables like weather)
- Non-Parametric Models:
- Gaussian Processes (Bayesian approach)
| Model | Forecast Performance (Test Data) | Model Fit (Training Data) | ||
|---|---|---|---|---|
| MAPE(%) ↓ | RMSE(kWh) ↓ | Log-Likelihood ↑ | AIC ↓ | |
| ARIMA(1, 0, 0) | 8.420 | 0.0224 | 2119 | -4232 |
| Auto ARIMA(2, 1, 1) | 16.41 | 0.0443 | 2136 | -4263 |
| Auto SARIMA(1, 1, 1)(2, 0, 2, 7) | 10.31 | 0.0285 | 2227 | -4509 |
| SARIMA(1, 1, 1)(2, 0, 2, 7) + Fourier for yearly variation | 21.27 | 0.0572 | 2248 | -4479 |
| SARIMA(1, 1, 1)(2, 0, 2, 7) + Weather | 4.870 | 0.0134 | 2311 | -4605 |
| Final Tuned SARIMA(1, 1, 2)(0, 1, 2, 7) + Weather | 2.900 | 0.0083 | 2304 | -4588 |
| GP (RBF & weekly periodic kernel) | 9.270 | 0.0005 | 43 | -73 |
| GP (RBF & weekly + annual periodic kernel) | 5.860 | 0.0002 | 67 | -116 |
