Investigation of Price-Feature Selection Algorithms for the Day-Ahead Electricity Markets

Abstract

This paper investigates three types of feature selection techniques such as relative importance using Linear Regression (LR), Multivariate Adaptive Regression Splines (MARS), and Random forest (RF) to reduce the forecasts error for the hourly spot price of the Iberian electricity markets. Two pricing datasets of durations three and six months were used to validate the performance of the model. Three different set of features (17, 4, 2) for three and six months duration were used in this study. These selected features were applied to the two-stage hybrid model such as ARIMA-GLM, ARIMA-SVM, and ARIMA- RF. Finally, three variables (or features) that are commonly matched were selected and tested. Considerable reduction in Mean Absolute Percentage Errors (MAPE) values were observed for both three and six-month datasets.