The Selection of Machine Learning Model and Its Hyperparameters Using Bayesian Optimization for Short-Term Wind Power Forecasting

Modern energy systems are often characterized by a significant share of the renewable energy sources installed capacity. Wind power is one of the most widespread renewable energy sources. Wind power forecasting is known to be a very effective measure for the wind farms integration into the power systems in terms of the short-term and ultra-short-term power end energy balances planning. This paper presents a technique for short-term wind power plants generation forecasting using a Bayesian optimization for model and its hyperparameters selection. The proposed approach provides for prior indirect selection of the most effective machine learning model that is expected to provide the best performance. The range of considered machine learning models includes regression tree, support vector machine, feed forward neural network, gaussian process regression and ensemble methods - random forest and gradient boosting over decision trees. The proposed approach provides a significant advance in prediction efficiency compared to manual selection of model and its hyperparameters halving normalized mean absolute error of a prediction.