Hybrid Deep Learning Framework for Short-Term Electricity Generation Forecasting in Türkiye Using Multi-Source Data

Abstract

The growing complexity of electricity generation, driven by the diversification of energy sources and the integration of renewables, makes accurate short-term
forecasting crucial for grid stability and energy security. This study proposes a deep learning-based hybrid forecasting model designed for Türkiye’s dynamic
energy landscape. Using hourly electricity production data from December 1, 2019, to March 1, 2025, sourced from the EPİAŞ Transparency Platform, the
model analyzes generation patterns across 17 different sources, including both fossil fuels and renewables. The proposed architecture combines Long ShortTerm Memory networks and Transformer models to effectively capture complex time-dependent relationships in electricity generation. To improve accuracy,
preprocessing techniques such as time-based interpolation, normalization, and principal component analysis were applied. Experimental results demonstrate
strong forecasting performance, achieving a mean absolute error of 589.50, a root mean squared error of 762.41, and a coefficient of determination (R2
) of 0.98017 for 1-hour ahead predictions, and an R2
of 0.87813 for 1-day ahead predictions. These findings underline the model’s potential to support operational
planning, market regulation, and policy-making processes, particularly in emerging economies with dynamic and heterogeneous energy infrastructures.

Cite this article as: Karamollaoğlu H. Hybrid deep learning framework for short-term electricity generation forecasting in Türkiye using multi-source data.
Turk J Electr Power Energy Syst. Published online October 20, 2025, doi 10.5152/tepes.2025.25029.