Problem Definition - Develop a time series model to Predict the power output of wind farm based on the weather condition in the site (1Hr prediction to 72Hrs. prediction) Build an application to recommend the Power Grid to suggest the best time to utilize the energy from wind farm
- The dataset provides us with the speed and direction of the wind recorded at a wind turbine.
- We calculate the sin and cos components of the wind using speed and direction.
- Time series prediction usies 4 different approaches to forecast the sin and cos components of the wind velocity. The four time seires models are Prophet, VAR, Holtwinters and GluonTS.
- Having obtained the predicted components of wind speed, we calculate estimated power generated by the wind turbines using MLP i.e. Neural Networks.
Wind Speed and direction affect each other. So we need a multivariate time seires model for better results. This purpose is served by VAR(Vector AutoRegression). For the Neural Network, we see that using sin and cos components of wind speed to compute the active power gives us less error than using wind speed and direction. This can be explained by the dependence but no direct correlation of wind speed and direction on Active Power.
- Facebook Prophet
- Vector AutoRegression (VAR)
- Exponential Smoothening Using HoltWinters
- GluonTS(MXNet) DeepAREstimator
- Multi-Layer perceptron (Neural Networks)
- Pandas
- Numpy
- Matplotlib
- Improve the existing time series model as the predictions are not very accurate.
- Use other DeepLearning techinques simultaneously for better results.
- Making better models using various other atmospheric data.
- Build a real-time interface predicting the power generated at various wind farms.
- Dataset - https://www.kaggle.com/berkerisen/wind-turbine-scada-dataset
- DeepAR Forecasting Tutorial - https://www.youtube.com/watch?v=g8UYGh0tlK0&list=WL&index=29&t=0s
- Note: The problem statement is taken from IBM Hackathon 2020