Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Automated Machine Learning Approach Applied to Nuclear Energy Generation Short-Term Forecasting
Submission Author:
Jorge Gustavo Sandoval Simão , SC
Co-Authors:
Jorge Gustavo Sandoval Simão, Viviana Mariani, Leandro dos Santos Coelho
Presenter: Jorge Gustavo Sandoval Simão
doi://10.26678/ABCM.COBEM2021.COB2021-0475
Abstract
Machine learning models applied to the time series forecasting problem are gaining interest among researchers and the industry and energy systems. This work evaluates automated machine learning, and feature selection (RreliefF) applied to time series forecasting for the United States nuclear power plants case study in a clustered dataset. This research evaluated forecasting results in terms of different performance metrics, where the AutoML (Automated Machine Learning) approach pipelined into the RreliefF algorithm presents promising results based on training and test datasets. To achieve this goal, nuclear power generation patterns were divided into two groups using clustering. Subsequently, the RReliefF algorithm is applied in each cluster, aiming to find the ideal number of features. To increase the precision of the model, a normalization was performed on the energy generation data of each cluster individually, and their skewness and kurtosis were measured and compared. The AutoML models that generated the best results using RreliefF are analyzed, and the metrics of R2 (Coefficient of Determination), MAE (Mean Absolute Error) and RMSLE (Root Mean Squared Logarithmic Error) are obtained. Despite a small difference, the number of features of each cluster shows that there is a difference in the generation of energy patterns, and that it is possible to generate accurate models of nuclear energy generation through the analysis of generation time series.
Keywords
Automated Machine Learning, Feature Selection, Nuclear Energy, Time Series, Forecasting
