E. Kouloumpris, A. Konstantinou, S. Karlos, G. Tsoumakas, and I. Vlahavas, “Short-term Load Forecasting With Clustered Hybrid Models Based On Hour Granularity”, In Proceedings of the 12th Hellenic Conference on Artificial Intelligence (SETN ’22), Corfu, Greece, 7-9 Sep 2022, Article 42, pp. 1–10, Association for Computing Machinery, 2022, DOI: https://doi.org/10.1145/3549737.3549783
Although the recent technological achievements have noticeable
impact on several aspects of daily life, more and more challenges
are raised in practice. As it concerns the Energy field, the need for
accurate predictions over time-dependent use cases of large scale
remains high. Deep learning approaches have already found great
acceptance in energy time-series signals, but there is still much
space for improvement. Contributing to the task of short-term
load forecasting we compose a hybrid method; first it exploits the
statistical profiling of input raw-signals validating them through
various complexity metrics; then a series of feature-engineering
processes are applied, before fitting a specified recurrent neural
network (RNN) architecture. During the first stage, we use time
series clustering to separate time periods in order to capture better
temporal patterns. We evaluate our approach using a public dataset
that regards the total load consumption of Spain, thus supporting
our assumptions about the benefits of leveraging hybrid models
for short-term load forecasting. The proposed method outperforms
other competitors, including a different RNN architecture and some
representative Machine Learning regressors.