Modeling and Prediction of the Energy Intensity Index by Provinces of Ecuador Using Neural Networks and Cluster Analysis with Machine Learning Algorithms

Abstract

Using a comparative experimental design, four neural network architectures—feedforward, LSTM, sequential feedforward, and robust LSTM—are evaluated to predict the provincial energy intensity index, based on panel data from 2018 to 2023 and key variables related to energy consumption, economic activity, and demographic dynamics. The sequential feedforward model demonstrated the highest predictive accuracy (R2 = 0.724), outperforming recurrent approaches in capturing temporal patterns, precision, and computational efficiency.

Clustering methods allow for the characterization of provinces’ behavior in terms of energy intensity, identifying groups with similar energy con sumption patterns relative to their economic output. Orellana, Pastaza, Santa Elena, and Zamora Chinchipe stand out for exhibiting distinct or atypical energy profiles, highlighting the usefulness of algorithms such as DBSCAN in detecting unique regional dynamics.
The results help identify regional differences in energy efficiency, providing an empirical basis for designing and implementing tailored energy policies. These insights are especially useful in varied geographic and socio-economic settings, where encouraging sustainable energy use and reducing disparities in distribution and consumption are key objectives.