A hybrid neuro-fuzzy inference system-based algorithm for time series forecasting applied to energy consumption prediction

The accuracy of the prediction of buildings' energy consumption is being tackled using existing artificial intelligence techniques. However, there is a lack of effort on the development of new techniques for solving that problem and, therefore, achieving higher performance, which is important for the efficient management of energy in many levels. This study addresses this gap by proposing a new hybrid machine learning algorithm that incorporates the adaptive neuro-fuzzy inference system model with a new version of the firefly algorithm denominated as the gender-difference firefly algorithm. We expanded the search space diversification to increase the accuracy on the prediction and adopted the autoregressive process in order to approximate the chaotic behavior of the consumption time series. A new layer, denominated as non-working time adaptation was also integrated so as to decrease the fast variability of the predictions during non-working periods of time. We have applied our algorithm for the consumption prediction on 1 h, 2 h and 3 h ahead horizons. We have obtained improvements on the MAPE and R coefficient when compared with state-of-the-art publications in both a private dataset from the Faculty of Chemistry, located in the city of Murcia, Spain and a public dataset of the consumption of a Retail building located in California, United States. We also show our method's performance in five more buildings. Our results demonstrate the robustness and the accuracy of our proposal when compared to the traditional adaptive neuro-fuzzy inference system models and also to the different predictive techniques implemented in several pieces of literature.