This repository contains a comprehensive study of short-term wind power forecasting using five models:
- SARIMAX (statistical model)
- Support Vector Regression (SVR)
- Random Forest
- XGBoost
- Long Short-Term Memory (LSTM)
The goal is to identify the most accurate and computationally efficient model for daily wind power prediction based on meteorological features.
├── data/
│ └── Turbine_Data.csv
├── results/
├── Wind_Power_Prediction.ipynb
├── README.md
└── LICENSE
We trained and tested all five models on a publicly available wind turbine dataset from Kaggle. The dataset includes Active Power, Wind Speed, Direction, Temperature, Pressure, and Humidity, recorded at 10-minute intervals and aggregated to daily level for day-ahead forecasting.
Key evaluation metrics:
- R² (Goodness of Fit)
- MAE (Mean Absolute Error)
- RMSE (Root Mean Squared Error)
- MAPE (Mean Absolute Percentage Error)
Random Forest provided the best fit to actual values, with SVR and SARIMAX also showing competitive performance. LSTM struggled due to limited granularity in the daily dataset.
- Random Forest outperformed all other models in accuracy and consistency.
- SVR performed well despite being linear.
- LSTM underperformed due to coarse daily aggregation and insufficient temporal data.
Random Forest proved to be the most reliable choice for short-term wind power forecasting with limited data. While LSTM has potential in high-frequency contexts, it is not suitable for daily aggregated datasets without significant volume.
- Use high-frequency (10-min) or hourly data for LSTM/GRU-based models
- Explore hybrid models combining statistical + ML/DL approaches
- Add probabilistic forecasting to assess prediction uncertainty
This project is licensed under the MIT License. See the LICENSE file for more details.