Short-term wind forecast using machine learning methods

More accurate and stable forecasting models are needed to reduce the system uncertainty caused by the variable nature of wind energy and to increase the revenues of the power plants by making more accurate production forecasts. For this purpose, two new hybrid models, both static and dynamic, and based on different machine learning algorithms, have been developed for the forecasting of short-term wind power in this study, unlike other studies. In the developed static model, a fixed training set containing data from all seasons was used. The Bayesian optimization algorithm was used to investigate the optimum hyper parameter values of Support vector regression (SVR), Decision tree (DT), Gaussian process regression (GPR), which are used as estimation algorithms in the static model. In order to increase the prediction performance of the optimized machine learning algorithms, ensemble learning models were created using bagging and stacking from ensemble learning algorithms based on combining multiple models. Long short-term memory (LSTM) neural networks, which gives successful results in time series analysis, was used in the developed dynamic model. Unlike the static model, the training dataset has a time-varying structure and was created with data from the previous 3 days (72 hours), 5 days (120 hours), 10 days (240 hours) of the test days. Neighborhood component analysis (NCA) algorithm was used to select the input parameter to be used with the historical wind power. Since it is a dynamic model, the network is updated with the actual power data of the previous time at each step. The developed models are applied for the 24-hour time horizon in the seasonal estimation of the wind power of a wind turbine in Turkey. The static bagging model was more successful than the dynamic NCA-LSTM model by taking the values of 11.045% Normalized Root of Mean Squares of Error (NRMSE), 4.880% Normalized Absolute Mean Error (NMAE) and 0.899 Coefficient of Determination (R^2). Although the dynamic model was more successful in times of sudden changes and fluctuations, the static bagging model generally gave results closer to the real power values since it was trained with much more data than each season.