A hybrid transformer–TCN–GRU based model for thermal load forecasting of a large university campus

Efficient energy management, cost reduction, and sustainability in campus operations require accurate forecasting of thermal loads in all seasons. This research addresses these objectives by deploying an advanced Hybrid Transformer deep-learning architecture specifically tailored to forecast long-term steam demand with exceptional accuracy and reliability. Using a comprehensive hourly steam demand dataset spanning eight years (2017–2024) collected from the Missouri Combined Cooling, Heating, and Power Plant and enriched with critical weather parameters (ambient air temperature, atmospheric pressure, wind speed, wind direction, relative humidity, and solar intensity). This study systematically explores complex dynamics influencing campus-wide steam consumption. The methodology involves meticulous data preprocessing, statistical analysis, exploratory analysis, and rigorous evaluation of the state-of-the-art-Transformer and Hybrid Transformer-TCN-GRU models against baseline models and comparative hybrid deep learning models. Performance metrics –Mean Absolute Error 5.04, Mean Squared Error 54.2, Root Mean Squared Error 7.36, Mean Absolute Percentage Error 0.0351, Mean Squared Logarithmic Error 0.0029, and Coefficient of Determination 0.96 – demonstrate the superior predictive capability of the Hybrid Transformer model, achieving higher accuracy, stability, and robustness compared to baseline and comparative models. The exceptional forecasting performance validates the Hybrid Transformer model's efficacy in handling intricate temporal dependencies, affirming its applicability as a central tool in campus energy management systems. This research contributes significantly to the integration of cutting-edge Transformer-based forecasting with detailed weather analysis, thereby advancing energy planning and sustainability initiatives in academic environments.