Univariate versus Multivariate Models for Short-term Electricity Load Forecasting

Abstract

Online short-term load forecasts are needed for efficient demand management on power systems. To model the load, univariate and multivariate forecast approaches were developed: while the first consider the load as a linear function of its time series, the other also takes in account the nonlinear effects of weather-related variables (mainly the air temperature). Despite the wide recent literature on multivariate models, some authors state that univariate ones are sufficient for short-term purposes, claiming that including temperature variables unnecessarily elevates the model complexity, putting parsimony and robustness at risk. In this study, we compare the forecasts produced, for real data, by several univariate and multivariate time series and neural network-based load curve models. We then use a nonparametric hypothesis test to compare the daily mean errors of the best forecaster of each kind and, so, verify if considering the air temperature leads to any statistically significant improvement in the forecasting.