On the Necessity of Exogenous Variables for Load, PV and Wind Day-Ahead Forecasts using Recurrent Neural Networks

In this paper we use publicly available time series data sets from the Global Energy Forecasting Competitions to evaluate the added value of exogenous, explanatory variables when forecasting wind, load and PV profiles. Two different auto-regressive models are built as well as one model that includes exogenous variables. All of models use recurrent neural networks (RNN) as their base architecture. The added value of exogenous variables is evaluated by comparing different accuracy metrics cross data set and cross model. The results show, that the autocorrelation of load and PV data sets produce reasonably good accuracies for auto-regressive predictions using RNNs, whereas wind production is far harder to forecast and the RNNs are not able to infer any suitable predictions using only a univariate time series.