Enhancing Power System Flexibility Using AI‐Based Forecasting Techniques: A Comparative Study

ABSTRACT This paper compares seven forecasting models for hourly electricity consumption in a commercial office building using data spanning 2024–2025. Models include XGBoost, LSTM, GRU, 1D‐CNN, SARIMA, Prophet, and Seasonal Naive baseline. Features encompass temporal indicators (hour, day of week, month), autoregressive lags (1, 2, 24, 168 h), and rolling statistics. Evaluation uses Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) on a 14‐day test set (336 samples) with rigorous hyperparameter tuning via GridSearchCV and TimeSeriesSplit cross‐validation. XGBoost achieves superior performance (MAE 6.29 kW, 3.5% MAPE) compared to GRU (10.95 kW), 1D‐CNN (11.86 kW), LSTM (14.98 kW), Seasonal Naive (16.15 kW), Prophet (35.72 kW), and SARIMA (48.16 kW). Paired t‐tests confirm statistical significance: XGBoost versus GRU () and versus Seasonal Naive (). Surprisingly, deep learning models underperformed gradient boosting despite theoretical sequence‐modeling advantages, attributed to modest sample size (17,016), rich feature engineering capturing 69.5% of variance through autoregressive features, and single‐hour forecasting horizon. Classical statistical models exhibited catastrophic failures, reflecting inadequate modeling of non‐stationary, non‐linear building consumption with multiple seasonal patterns. Results demonstrate that for structured tabular time series with comprehensive feature engineering, gradient boosting substantially outperforms sequential neural architectures and classical statistical methods. The findings enable high‐confidence building energy management decisions (HVAC pre‐conditioning, demand response) with ±6.3 kW prediction accuracy. Code and reproducibility documentation are available at https://github.com/SaadHayat91/BEJ-Electricity-Forecasting .