Explainable Transformer-Based Energy Consumption Forecasting for Green Smart Campuses Using Multi-Modal IoT Data
Year: 2026
Type: conference-paper
Abstract: Accurate and interpretable energy consumption forecasting is essential for efficient operation and sustainability of smart campuses. The increasing deployment of Internet of Things (IoT) infrastructure enables the collection of multi-modal data capturing energy usage, occupancy behavior, environmental conditions, and renewable energy generation. However, effectively exploiting such heterogeneous data while maintaining model transparency remains a significant challenge. This paper proposes an explainable Transformer-based energy consumption prediction framework using a Temporal Fusion Transformer (TFT) architecture for green smart campus environments. The proposed model integrates multi-modal IoT data, including energy, occupancy, weather, solar generation, and sustainability indicators, to capture both short-term dynamics and long-range temporal dependencies. Experimental evaluation on a real-world smart campus dataset demonstrates that the proposed model significantly outperforms baseline deep learning approaches, achieving an Root Mean Squared Error (RMSE) of 72.74 kWh and Mean Absolute Error (MAE) of 61.59 kWh, compared to RMSE values above 179 kWh for LSTM and GRU models. In addition, permutation-based feature importance and ablation studies provide transparent insights into the dominant factors influencing energy consumption. These results confirm that the proposed framework delivers accurate, robust, and explainable energy forecasting suitable for smart campus energy management and sustainability-oriented decision support.
Institution: DY Patil Deemed to be University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204804722