Wind speed forecasting approach using conformal prediction and feature importance selection
Cesar Vinicius Zuege, Stéfano Frizzo Stefenon, Cristina Keiko Yamaguchi, Viviana Cocco Mariani, Gabriel Villarrubia González, Leandro dos Santos Coelho
Abstract
Wind energy is a rising renewable energy source that plays an important role in the transition to a more sustainable energy system. Variation in wind power generation is one of the main challenges facing this energy source. Wind forecasting approaches are essential for planning and operating wind farms, but are complex due to the dynamic nature of the wind and the influence of local factors. This paper evaluates short-term forecasting of time series with a measure of uncertainty associated with wind speeds. The proposed method considers the conformal prediction approach and, based on Shapley values, uses optimal selection of features given their importance. Furthermore, a Bayesian Optimization with Tree-structured Parzen Estimators (BO-TPE) will be used to tune the hyperparameters of the models. The results showed that using Variational Mode Decomposition (VMD) allied with Singular Spectrum Analysis (SSA) to feed into a conformal prediction model improved the performance of the model. Taking into account a Beutenberg data set, Germany, the best model was Light Gradient Boosting Machine (LGBM)-VMD-SSA without partial fit, resulting in a root mean squared error (RMSE) criterion of 0.25031, coverage measure of 94.4%, and width measure of 1.008. When considering a dataset from Limoeiro, Brazil, the best model was also LGBM-VMD-SSA without partial fit, resulting in an RMSE of 0.21597, a coverage of 90.3%, and a width of 0.678. SHapley Additive exPlanations (SHAP) bring explainability to the model results. The models proposed in this study, hypertuned by BO-TPE with interpretable results based on SHAP, can be useful in predicting wind speed and power generation. • Improvement of energy planning based on time series forecasting in wind farms. • Variational mode decomposition with singular spectrum analysis for data preprocessing. • Probabilistic forecasting using conformal prediction. • Shapley additive explanation for explainable artificial intelligence.