Benchmarking Seasonality‐Based Deep Learning Models for Short‐Term Load Forecasting: Prophet, N‐BEATS, N‐HiTS and DLinear Across European Power Grids

ABSTRACT The rapid integration of intermittent renewable energy sources necessitates high‐precision short‐term load forecasting (STLF) to ensure grid stability and efficient supply‐demand balancing. This paper presents a rigorous comparative evaluation of four seasonality‐based state‐of‐the‐art forecasting architectures—Prophet, N‐BEATS, N‐HiTS and DLinear—benchmarked against four conventional machine‐learning models that dominate the STLF literature (random forest, support vector regression, XGBoost and an LSTM/GRU recurrent network), applied to 15‐min resolution load data from six European countries (Austria, Belgium, Germany, Hungary, Luxembourg and the Netherlands). The study utilizes a multi‐year dataset (2015–2020) and evaluates performance during the COVID‐19‐driven demand shifts of 2020, serving as a stress test for model robustness under extreme distributional changes typical of modern power systems. A dedicated calendar‐feature ablation and an event‐stratified error decomposition further quantify how weekends, public holidays and irregular events influence comparative accuracy. Results identify a clear performance hierarchy: the statistical‐heuristic Prophet model fails on smaller, volatile grids; the conventional machine‐learning baselines improve substantially upon it yet remain well short of the decomposition‐based architectures and inherit their fragility under distributional shift; and N‐BEATS exhibits significant generalization gaps between validation and test phases. Conversely, N‐HiTS and DLinear demonstrate exceptional resilience, consistently achieving and MAPE below across all grid scales and remaining below MAPE even on public holidays and during the pandemic onset. Crucially, the minimalist DLinear architecture—utilizing moving‐average decomposition—matches or exceeds the performance of far more complex models. These findings offer system operators actionable insights into selecting robust, computationally efficient forecasting tools that are essential for managing the increased variability and uncertainty inherent in high‐renewable penetration power systems.