Wind power forecasting with machine learning methods

The negative effects of climate change, finite reserves of fossil fuels and reduction of foreign dependence on energy have increased the importance of renewable energy. Fossil fuel reserves' shortage and incentives, increased the rate of renewable resources in our country's installed power capacity. Installed capacity of wind power plants has exceeded 10 GW as of 2021. While wind being a renewable resource is an advantage, intermittent characteristic causes problems for system operator. Wind power forecasting is very important in order to solve these problems and to meet the needs of the electricity market and wind farms. In this study, for wind forecasting, the data obtained from turbine anemometer in one year time horizon of 10 minutes. Models are designed with univariate long short term memory, multivariate long short term memory, multilayer perceptron, extreme gradient boosting, convolutional neural networks-long short term memory, gated recurrent unit and random forest algorithms. Outliers in the data are cleaned with the isolation forest algorithm. Hyperparameter optimization of the models, designed with machine learning algorithms for a selected month from each season (January, April, July, November) was made with Bayesian optimization. The success of the models was measured by mean absolute error, root mean square error and determination coefficient. Among the designed models, for January, April and November, the multivariate long short term memory model gave the most succesful results with 0.926, 0.961 and 0.946 determination coefficient values, respectively. For July, extreme gradient boosting model gave the most successful result with 0.916 value of determination coeffcient. In this study, for wind forecasting, the data obtained from turbine anemometer in one year time horizon of 10 minutes. Models are designed with univariate long short term memory, multivariate long short term memory, multilayer perceptron, extreme gradient boosting, convolutional neural networks-long short term memory, gated recurrent unit and random forest algorithms. Outliers in the data are cleaned with the isolation forest algorithm. Hyperparameter optimization of the models, designed with machine learning algorithms for a selected month from each season (January, April, July, November) was made with Bayesian optimization. The success of the models was measured by mean absolute error, root mean square error and determination coefficient. Among the designed models, for January, April and November, the multivariate long short term memory model gave the most succesful results with 0.926, 0.961 and 0.946 determination coefficient values, respectively. For July, extreme gradient boosting model gave the most successful result with 0.916 value of determination coeffcient.