As energy systems become increasingly data-driven, the ability to accurately forecast electricity consumption has emerged as a strategic challenge for both industry and national infrastructure planning.
Today, we proudly highlight the work of 𝘿𝙧. 𝘿𝘼𝙇𝙄𝙇 𝙃𝘼𝘿𝙅𝙊𝙐𝙏, whose scientific publication in the prestigious Expert Systems with Applications, published by [Elsevier], contributes to advancing Artificial Intelligence applications in energy forecasting and multiple time-series analysis.
Co-authored with Abderrazak Sebaa, José F Torres, and Francisco Martinez-Alvarez, the study introduces a robust forecasting framework designed for medium-term electricity consumption prediction within the Algerian economic sector.
Entitled “Electricity consumption forecasting with outliers handling based on clustering and deep learning with application to the Algerian market”, the research addresses one of the major challenges in modern AI systems: forecasting multiple heterogeneous time series generated by large-scale electricity consumers.
Using real-world data collected from more than 2000 economic-sector consumers over a 14-year period in Bejaia and provided by Sonelgaz, the study focuses particularly on High Voltage type A (HVA) consumers, whose energy demand represents a significant share of Algeria’s electricity consumption.
The proposed methodology is structured around three complementary stages:
– intelligent preprocessing and outlier management through adapted Robust Exponential and Holt-Winters Smoothing techniques;
– customer profiling using K-Means and K-Shape clustering algorithms to identify structural consumption similarities;
– deep learning forecasting using GRU (Gated Recurrent Unit) neural networks optimized through Bayesian hyper-parameter tuning.
Beyond the methodological contribution, the framework demonstrates strong operational applicability for electricity distribution companies. By enabling more accurate demand anticipation and customer consumption analysis, the proposed solution contributes to reducing non-technical losses, optimizing field interventions, and improving energy management strategies.
The experimental results demonstrate remarkable predictive performance. According to the MAPE metric, the proposed framework achieved an error rate as low as 2.04%, while 87% of customers exhibited considerably low forecasting errors. The study also showed the superiority of the K-Shape-based clustering approach compared with several established forecasting methods, including SARIMA, SGM, LSTM, TCN, and ensemble models.
Importantly, this research represents one of the first advanced studies dedicated to Algerian electricity consumption forecasting using clustering and deep learning techniques in a multiple time-series context. The work therefore contributes not only to the scientific literature in Artificial Intelligence and energy forecasting, but also to the development of practical AI-driven solutions for real-world industrial systems.
Through contributions such as this, ESTIN continues to foster high-level applied research at the intersection of Artificial Intelligence, Data Science, and strategic technological innovation.