Theoretical and Empirical Analysis of ReliefF and RReliefF
Machine Learning
Hi-index | 0.00 |
We present a new approach for feature selection for 5-minute ahead electricity load forecasting, using load data of the previous 4 weeks. An initial set of features is selected based on the weekly and daily load patterns. This set is then finalized by assessing the merit of the candidate features using feature selection algorithms that capture linear and nonlinear dependencies. The best secondary feature selector was a combination of two ranking-based algorithms, RReliefF and mutual information, achieving an error MAPE of 0.282% with backpropagation neural networks and 0.283% with linear regression and using less than 1% of the original features. Our approach was also considerably more accurate than the state-of-the-art Holt-Winters exponential smoothing methods, the industry model and a number of other baselines.