A Reproducible Comparison of ARIMA and LSTM for Hourly Electricity Load Forecasting under a Rolling One-Step-Ahead Protocol

Short-horizon forecasting of electricity load is a core task in power-system operation, and it is often unclear whether modern recurrent neural networks justify their cost over classical statistical models. This work compares an ARIMA model and an LSTM network on hourly electricity-consumption data from the PJM Interconnection (AEP zone, 2004-2018, public-domain Kaggle dataset). Both models are evaluated under an operational rolling one-step-ahead protocol in which each model has access to the full actual history up to the forecast origin and predicts one hour ahead, after which the realised value is fed back. Under this protocol the two approaches are statistically indistinguishable in five-fold time-series cross-validation (RMSE about 1500 MW, R-squared about 0.79 for both), while on the held-out test set ARIMA(2,0,2) is modestly better than the LSTM (RMSE 2320 vs 2468 MW; MAE 1245 vs 1636 MW; R-squared 0.813 vs 0.788). An error-distribution analysis shows ARIMA produces fewer large misses, and the LSTM begins to overfit within a few epochs. ARIMA also trains about 11.5x faster and remains interpretable. A transferability test on a previously unseen utility (Commonwealth Edison) preserves accuracy for both models (R-squared about 0.83 / 0.81), indicating no dataset-specific overfitting. The experiments are wrapped in a client-server prototype (FastAPI + React) that loads arbitrary hourly series, auto-suggests hyperparameters, trains both models and visualises the comparison. The main practical conclusion is that at a one-hour horizon with feedback, a well-specified linear model is a strong, cheaper baseline that a single-layer LSTM does not surpass.