Energy consumption forecasting in large buildings, particularly hotels, is critical for efficient energy management, operational cost reduction, and decarbonization. Their energy demand is highly dynamic, influenced by fluctuating occupancy, weather conditions, and building thermal behavior. This study benchmarks a hierarchy of data-driven models — from simple univariate approaches to complex multivariate ones — integrating occupancy, weather, and calendar features, as well as low-fidelity multi-physics simulation data for a large hotel tower in Madrid, Spain. Through systematic feature selection and ablation analysis, we identify the most influential predictors and evaluate several machine learning algorithms, including Gradient Boosting, LightGBM, Multi-Layer Perceptron, and Long Short-Term Memory. The best-performing model (LightGBM with simulated thermal data) achieved a mean absolute percentage error (MAPE) of 4 . 27 ± 3 . 69 % , demonstrating the added value of combining physical insight with data-driven forecasting. Yet, simpler models relying solely on historical data achieved comparable accuracy, underscoring that robust predictions can be obtained even under constrained data conditions. These findings advance the integration of artificial intelligence and building physics for smart energy management, providing actionable guidance to practitioners and promoting the broader deployment of forecasting tools that support the sustainable operation and decarbonization of the hospitality sector. • Benchmark of data-driven models for short-term hotel electricity forecasting. • Integrate occupancy, weather, calendar, and simulation data to capture building dynamics. • Feature-ablation analysis to reveal top predictors under data constraints. • Tree models (LightGBM, GB) reach <5% MAPE, rivaling more complex neural networks. • Provide guidelines for efficient forecasting models in smart, low-carbon hotels.
Flórez et al. (Thu,) studied this question.