Current industrial practice of predicting the power generation heavily relies on a method known as the binning method which is recommended by International Electrotechnical Commission (IEC). This is a simple approach where the main predictor, i.e., wind speed, is divided into discrete bins. The value to be used for representing the power output for a given bin is average for all the data points in each of those bins.
Above mentioned method is a simple method which takes into consideration only a single predictor namely “Wind speed”. There are other factors which may significantly impact the power generation but are ignored by this method.
The data for aforementioned analysis has been collected from a wide array of sensors installed on the meteorological mast in the wind farm. Data collection period is one year, so as to factor in the seasonal changes as well. The data has been averaged over 10 minute intervals for ease of analysis. We have done some data exploration to understand the kind of data we are dealing with.
This project aims to try to factor in the predictors that may impact the power output and see their correlation. We have tried some popular learning methods like Gaussian process regression, SVM, CART, Random Forest along with methods like AdaBoost to enhance prediction. We also have used basic methods like linear regression, sparse models like ridge and lasso. In this era of deep learning, we have avoided using Neural Network because the method is kind of a “black box” and is difficult to explain to a layman. We have used python, a popular programming language across the industries, to carry out the analysis.
Some of the methods use sample from the data rather than the whole data for training. We have used training and validation set, to get a proper idea of accuracy. Cross validation has been applied to SVM and Random Forest for “hyper-parameter” tuning.
Root Mean Square Error has been used as a measure for accuracy. Even though, CART being a simpler method as compared to other algorithms like SVM and GPR, it gives the least RMSE among all the methods used for prediction. Might be because of how the data is inherently structured. Finally we compare the RMSE of all the models and use the best among those methods for final prediction on test dataset.
The project report can be found here
A short presentation summarizing the project can be found here
Code for the whole project can be found here