Deep Learning Benchmark for National Electricity Consumption Forecasting: Architecture Comparison and Energy Security Implications for Türkiye

Accurate forecasting of hourly electricity consumption is critical for smart grid management, energy market operations, national policy planning, and—particularly for import-dependent economies such as Türkiye—energy security. This study presents, to the best of the authors’ knowledge, the first systematic benchmark of four state-of-the-art time series architectures—TimesNet, PatchTST, iTransformer, and Temporal Fusion Transformer (TFT)—conducted specifically on a national-scale Turkish multivariate energy dataset from the Energy Exchange Istanbul (EPİAŞ), covering 72,322 hourly observations across 15 generation, consumption, and market-clearing price variables from January 2018 to April 2026. While benchmark studies of Transformer-based architectures exist on general time-series datasets, no prior work has applied this specific combination of architectures to the EPİAŞ dataset under unified experimental conditions with an explicit energy-security interpretation. All models were trained under standardized preprocessing (StandardScaler), a 24 h lookback window, and systematic hyperparameter optimization. Experimental results demonstrate that iTransformer achieves the best predictive performance (MAE = 521.34 MWh, RMSE = 748.12 MWh, R2 = 0.9881, MAPE = 1.34%), followed by TFT (R2 = 0.9863) and PatchTST (R2 = 0.9844). TimesNet, while the most computationally efficient, achieves an R2 of 0.9791. Beyond predictive benchmarking, this study situates the findings within Türkiye’s energy security agenda: the dataset captures fossil fuel dependency, the growing share of domestic renewables, and market-clearing price dynamics shaped by geopolitical shocks, including the Russo–Ukrainian war and evolving EU–Türkiye energy relations. Comprehensive analysis of model architectures, attention mechanisms, temporal feature importance, and computational efficiency is provided. These findings establish a rigorous baseline for deploying modern sequence models in large-scale, real-time national energy forecasting systems that serve both market-efficiency and strategic-energy-autonomy objectives. The results specifically highlight how high-fidelity forecasting can serve as a risk-mitigation tool against geopolitical supply disruptions by quantifying the impact of domestic renewable integration.