Comparative Evaluation of Deep Learning Architectures for Electricity Demand Forecasting

Source abstract

This study investigates univariate multi-horizon forecasting of national electricity demand as a controlled benchmark for settings where exogenous drivers (e.g., weather and calendar variables) are unavailable or uncertain, through a comparative evaluation of representative deep learning architectures. The examined models include the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, a Temporal Convolutional Network (TCN), and the feed-forward Neural Basis Expansion Analysis for Time Series (N-BEATS) framework. All models are trained and evaluated within a unified experimental setup based on a univariate daily time series of Finnish national electricity demand covering the period from 2016 up to 2021, enabling a controlled assessment of architectural capabilities when relying solely on historical demand. Using a common preprocessing pipeline and a chronological train–validation–test split, forecasts are generated for short-, medium-, and long-term intervals (30, 90, and 365 days), and predictive performance is assessed using the Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The experimental results show that N-BEATS achieves the lowest RMSE across all considered horizons in the test set, while the GRU architecture attains the smallest MAE at the longest horizon and exhibits consistently strong performance overall. These findings highlight the complementary strengths of recurrent and feed-forward deep learning paradigms for modelling nonlinear structure and long-range dynamics in electricity demand time series, and provide quantitative evidence to support horizon-aware architecture selection in national electricity demand forecasting and related applied modelling contexts.