Intermediate Short-Term Wind Power Forecasting Using LSTM

Key Features:
1. Developed LSTM models for forecasting wind power generation 3 hours and 8 hours ahead.
2. Optimized model architectures for different window sizes (1-20 hours) and hyperparameters.
3. Implemented 1D convolution for longer sequence models, improving performance by about 10%.

Technical Details:
1. Utilized wind forecast and load forecast data from MISO (Midcontinent Independent System Operator).
2. Applied various LSTM architectures, including stateful LSTM and 1D convolution for temporal data.
3. Experimented with different activation functions, including 'relu', 'linear', and 'tanh' combinations.
4. Employed 'Adam' optimizer for model training.
5. Conducted uncertainty analysis by introducing artificial noise to input data.

Impact:
1. Achieved high accuracy in predictions:
- Best 3-hour forecast model (4-hour window): MAPE of 9.77%, RMSE of 0.0559
- Best 8-hour forecast model (20-hour window): MAPE of 11.40%, RMSE of 0.0564
2. Demonstrated robustness of models under various levels of measurement error through uncertainty analysis.
3. Developed models that outperform or are comparable to existing wind power forecasting systems like WPMS and AWPPS.
4. Potential to improve real-time grid operations, load dispatch planning, and overall energy management efficiency.

This project showcases advanced time series forecasting techniques applied to the critical domain of wind power prediction, with potential applications in grid management and energy planning.