Energy Consumption Forecasting in a University Office by Artificial Intelligence Techniques: An Analysis of the Exogenous Data Effect on the Modeling

The forecasting of building energy consumption remains a challenging task because of the
intricate management of the relevant parameters that can influence the performance of models. Due
to the powerful capability of artificial intelligence (AI) in forecasting problems, it is deemed to be
highly effective in this domain. However, achieving accurate predictions requires the extraction of
meaningful historical knowledge from various features. Given that the exogenous data may affect
the energy consumption forecasting model’s accuracy, we propose an approach to study the importance
of data and selecting optimum time lags to obtain a high-performance machine learningbased
model, while reducing its complexity. Regarding energy consumption forecasting, multilayer
perceptron-based nonlinear autoregressive with exogenous inputs (NARX), long short-term
memory (LSTM), gated recurrent unit (GRU), decision tree, and XGboost models are utilized. The
best model performance is achieved by LSTM and GRU with a root mean square error of 0.23. An
analysis by the Diebold–Mariano method is also presented, to compare the prediction accuracy of
the models. In order to measure the association of feature data on modeling, the “model reliance”
method is implemented. The proposed approach shows promising results to obtain a well-performing
model. The obtained results are qualitatively reported and discussed.