A Comparative Study of Machine Learning, Deep Learning, and Foundation Models for Short-Term Electricity Load Forecasting

Abstract

Short-term electricity load forecasting is a key task for smart-grid operation, industrial energy management, and predictive planning. This paper presents a reproducible comparative study of classical, machine-learning, deep-learning, and foundation-model approaches for hourly load forecasting. The benchmark includes SeasonalNaive, SARIMA, XGBoost, DLinear, PatchTST, iTransformer, Chronos-Bolt-Small, and TimesFM 2.5. The primary dataset is ECL with aggregate electricity load as the target; ETTh1/HUFL is used as a secondary industrial benchmark. Experiments cover 24-, 96-, and 168-hour horizons with a 336 -hour context window. On ECL, the supervised neural models - iTransformer, DLinear, and PatchTST-achieve mean MAE roughly an order of magnitude lower than Chronos-Bolt-Small, TimesFM 2.5, and a grid-tuned SARIMA at every evaluated horizon. A controlled training-budget ablation shows that this gap is not a training-budget artefact: the supervised models converge within approximately 1000 optimizer steps and gain less than 4% in MAE between 500 and 5000 steps. On ETTh1/HUFL the ordering is more compressed: Chronos-Bolt is best at $\mathbf{H} \boldsymbol{=} \mathbf{2 4}$, while DLinear is best at $\mathbf{H} \boldsymbol{=} \mathbf{9 6}$ and $\mathbf{H} \boldsymbol{=} \mathbf{1 6 8}$, with foundation models within a few percent. Diebold-Mariano tests on seed-averaged per-window errors confirm the leader ordering is statistically significant. The study also reports training time, single-window inference latency, and peak GPU memory, showing that foundation models remain attractive cold-start tools but, under the evaluated zero-shot configuration, do not universally replace properly trained supervised baselines.