Benchmarking Deep Learning Architectures for 24-Hour Energy Forecasting in Smart Buildings Using Real-World IoT Data

Abstract

Accurate 24-hour energy forecasting in smart buildings remains a challenging real-world problem due to the highly non-linear, dynamic, and heterogeneous nature of IoT sensor data. This study extensively expands upon previous research that demonstrated the effectiveness of Bi-LSTM as a univariate predictor. To address existing gaps, we make two primary contributions: (1) developing a multivariate forecasting framework that incorporates eight diverse IoT sensor streams, including HVAC (heat pump) energy consumption and indoor environmental comfort metrics, and (2) benchmarking advanced attention-based architectures, such as the Temporal Fusion Transformer (TFT), optimized for interpretable, high-performance multidimensional time-series forecasting. Using real-world datasets from a multifunctional smart building in France (Pôle Culturel), experiments were conducted over approximately four months (October 23, 2024, to February 27, 2025) at 10-minute intervals. The preprocessing pipeline includes timestamp harmonization, Akima interpolation for non-linear data gaps, outlier correction, and feature scaling. Using MAE, RMSE, MAPE, and R² as evaluation metrics, the experiments compare a multivariate Bi-LSTM against standard models such as LSTM, GRU, and the TFT. The results indicate that the multivariate Bi-LSTM (R² = 0.9656) significantly reduces errors compared to univariate approaches, accurately capturing daily trends and peak loads. Furthermore, while the TFT effectively models complex multivariate dependencies (R² = 0.8130), it demonstrates that attention-based models require specific architectural tuning for medium-scale IoT datasets, providing practical guidance for sustainable, data-driven energy management.