Short-term forecasting of wind power production using machine learning and deep learning methods

The primary purpose of this thesis is, by using relating data, to make short-term wind power prediction with the aid of Deep Learning and Machine Learning technology. The main methods of the research presented in these studies have implemented in the Regression Methods and Artificial Neural Network Method. The related turbines are located in the wind farm on Harmanlık in Bursa. It had established in the middle of 2015, and the wind park consists of 16 number Vestas 112 wind turbines models. Power generation capacity is 52.8 MW as annual. In this thesis is going to focus on only eight turbines. Machine Learning has been on the predictable and applicable a model. This study is constituted the prediction algorithm with various regression methods which are Multi Linear Regression, Support Vector Regression, K-Nearest Neighbors, and Decision Tree Regression. These prediction algorithms have been to improve a software application in Python programs. Furthermore, Deep Learning provides a significant part in the implement of optimization. These method has also been to establish the optimum configuration in a software application Python programs to reach the prediction of wind power. These regression methods and Artificial Neural Network has used on various data containing the wind speed, the temperature, and past-wind power production. The processed data was received during years from 01.09.2016 to 31.08.2018 by Borusan Enbw energy company and the Borusan Enbw meteorological sensor in the wind farm. Based on the comparison of daily data has indicated that the artificial neural network and the whole regression analyses might obtain performances, that is the regular (24 hours) forecasting for a single turbine, whether which prediction results are equal or even better compared to each other. After all simulation, it will decide that the best method to adopt for this the prediction of wind power within among methods. All forecast result generated with these identified methods. All considered models are using wind speed of the location and the temperature of the area from sensors. Moreover, it has been to apply the most critical data which is past production data from 01.09.2016 to 31.08.2018 as an hourly. After that, it has implemented two analysis. One of them will compare between the prediction results, which obtains only wind speed and, the prediction results, which receive both wind speed and the temperature to each other for one turbine by using the Regression Methods and Artificial Neural Network. According to these solutions, it was decided to use two input parameters, which are the wind speed and the temperature. It was implemented the related inputs for five methods. It has a presentation of graphs detailing how well the methods ("Multi Linear Regression," "Support Vector Regression," "K-Nearest Neighbors," "Decision Tree Regression" and "Artificial Neural Network") perform in the studies. Furthermore, the second analysis has compared to the accuracies for eight turbines by using target data and prediction data. These results are also briefly explain in this thesis. It is shown the various prediction results and their relative performance by measuring MSE, MAE, RMSE.