Ensemble forecaster based on the combination of time-frequency analysis and machine learning strategies for very short-term wind speed prediction

Abstract

• A time series-frequency forecaster is proposed for wind speed prediction. • The hyperparameters of different time series-frequency analysis were optimised. • The forecaster combines time series-frequency analysis and machine learning algorithms. • Final forecaster makes more accurate predictions than the benchmark and other models from the literature. Based on the predictions of fossil fuels depletion in the following years, as well as their negative impact due to generated exhaust fumes, eco-friendly generators, and more specifically wind generators, have arisen as a solution for the electric demand challenge. Wind energy consists in extracting energy from wind speed, and because of the uncertain and intermittent behaviour of this meteorological parameter, wind turbines output power cannot be optimally exploited. Although the vast majority of the research in wind speed forecasting field has consisted in the purpose of novel algorithms, these studies have not made a pre-processing step of the data in order to try to extract the maximum information from databases. Therefore, the goal of this paper consists in analysing whether the combination of time-frequency decomposition of wind speed data with different machine learning algorithms can increase the accuracy of current wind speed predictions for 10 min ahead. Obtained error metrics demonstrated that the deviation of developed wind speed forecaster was lower than 0.1% in 62% of the validation database. In addition, the root mean square error of the final forecaster was 0.34 m/s. This means an accuracy increase of 51.5% if the result is compared with benchmark model's results.