Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
An improved ensemble learning model for multi-step ahead wind power generation forecasting.
Submission Author:
Matheus Henrique Dal Molin Ribeiro , PR
Co-Authors:
Matheus Henrique Dal Molin Ribeiro, Ramon Gomes da Silva, Sinvaldo Rodrigues Moreno, Viviana Mariani, Leandro dos Santos Coelho
Presenter: Sinvaldo Rodrigues Moreno
doi://10.26678/ABCM.COBEM2021.COB2021-0166
Abstract
Wind energy is a renewable, non-polluting, and clean energy resource, which the generation increase over the last years in Brazil. Then, forecasting wind power generation is key to developing correct strategic planning in energy power distribution. Forecasting wind power energy is a challenge. However, due to the uncertainty associated with wind speed, it is a challenge. Therefore, this study evaluates bootstrap aggregation efficiency (bagging) combined with a stacking ensemble learning model for short and medium-term (one up to twelve-hours-ahead) forecasting of wind power generation a wind turbine for a wind farm located in Parazinho, Brazil. In this approach, we aim to identify the suitable number of samples (10, 30, 50, or 100) obtained through bootstrap to compose the bagging model; and obtain the best strategy used in the bagging model, average or median. The proposed framework is composed of three steps, (i) The Seasonal and Trend decomposition using Loess (STL) is applied in the data; (ii) the moving block bootstrapping is used for the remainder component, and next the original signal is reconstructed by summing up the bootstrapped remainder, trend, and seasonal components; and (iii) the stacking ensemble learning is applied in each sample to obtain the forecasting results. Next, the final forecasting of all bootstrap samples is aggregated by average or median. The stacking ensemble learning base models are Gaussian processes and Support Vector Regression with linear kernel, k-Nearest Neighbor, and Random Forest. The meta-model is the ridge regression. The forecasting accuracy is evaluated through the root mean squared error, mean absolute error, and Theil’s U index of inequality. In a broader perspective, in 83.33% of the comparisons, the stacking combined with bagging ensemble have better accuracy than the stacking ensemble learning model regarding the performance measures. The results suggest that for one-hour-ahead forecasting wind power generation, the stacking ensemble learning achieves forecasting errors lower than the combination of stacking with bagging ensemble approach according to all performance criteria. Also, these two approaches have competitive results concerning the forecasting horizons of two and three-hours-ahead. When dealing with the number of bootstrap samples, better forecasting results regarding the performance measures are achieved when 30 bootstrap samples are considered in the ensemble structure, followed by 100, 50, and 10. Regarding the aggregation strategy, the average and median approaches provide similar forecasting accuracy regarding the adopted criteria.
Keywords
Wind Power Generation, Time series forecasting, Ensemble Learning, machine learning, Artificial Intelligence
DOWNLOAD PDF VIEW PRESENTATION
