Feature Selection for Neural Network-Based Interval Forecasting of Electricity Demand Data

We consider feature selection for interval forecasting of time series data. In particular, we study feature selection for LUBEX, a neural-network based approach for computing prediction intervals and its application for predicting future electricity demands from a time series of previous demands. Our results show that the mutual information and correlation-based feature selection methods are able to select a small set of lag variables that when used with LUBEX construct valid and stable prediction intervals (coverage probability of 97.44% and 96.68%, respectively, for confidence level of 90%). In contrast, the popular partial autocorrelation feature selection method fails to do this (coverage probability of 69.69%). Our evaluation was conducted using one year of half-hourly Australian electricity demand data.