Influência do tratamento de dados no desempenho de modelos de previsão da geração eólica usando informações de restrição de potência pelo Operador Nacional do Sistema Elétrico

Abstract

The transition to renewable energy sources highlights the need for accurate predictions, especially for wind power generation, where its intermittent nature challenges the reliability of the electrical system. In response to this factor, the restriction of renewable energy production, applied by the National Electric System Operator (ONS), has grown and is introduced as another variable in wind power generation. Accurate forecasting methods can contribute to reduce operating costs and increase the reliability of the electrical system. For this, various steps are involved, such as pre-processing, predictive modeling, cross-validation, and results analysis. In this study, the effect of this set of practices on wind power forecasting was investigated, with emphasis on pre-processing, considering restriction scenarios based on data extracted from the ONS for the "SERRA DE SANTANA 3" wind complex in Rio Grande do Norte, with an installed capacity of 108 MWp. After normalization by installed capacity, supervised learning techniques and algorithms such as Least Absolute Shrinkage Operator (LASSO), Ridge, Light Gradient Boosting Machine (LGBM), and Support Vector Regressor (SVR) with Bayesian optimization were applied for 1-hour ahead prediction. Different data modeling techniques, including wind power curtailment as an external variable, were applied to train these algorithms. For performance comparison, reference models using only energy generation data and the persistence method were used. Based on the results, it can be concluded that the generation restriction, being a sparse property, had little importance for the models compared to other attributes. The SVR method, using pre-processing techniques, stood out the most in terms of improvement over reference models. The performance analysis of the methods was based on the metrics Normalized Mean Absolute Error (NMAE), NRSME, and R2 in percentage terms, resulting in SVR with the proposed pre-processing as the best method with NMAE of 5.30%, NRMSE of 7.42%, and R2 of 90.70%. The main contribution of the study was the evaluation of the wind power forecasting using different machine learning methods under different contexts and considering the growth of wind power curtailment imposed by the ONS.