Hybrid Deep Learning and Ensemble Approach for HVAC Energy Forecasting: A GRU + Random Forest Framework

Heating, Ventilation, and Air Conditioning (HVAC) systems are among the most energy-intensive components of modern buildings, responsible for nearly 40% of global building energy consumption. An accurate prognostication of HVAC energy consumption is consequently imperative for formulating strategies aimed at enhancing efficiency and minimizing expenses. Conventional machine learning (ML) frameworks, such as Random Forest (RF), demonstrate commendable performance yet encounter difficulties in identifying sequential patterns within time-series data. On the other hand, deep learning (DL) designs, such the Gated Recurrent Unit (GRU), are good at capturing temporal dependencies, but they often need a lot of computing power and are prone to overfitting. This paper presents a new hybrid forecasting model that combines the Gated Recurrent Unit (GRU) and Random Forest (RF) methods to take advantage of the best features of both. A two-stage process was used with the Oak Ridge National Laboratory (ORNL) FRP-2 multizone building dataset. In Stage 1, GRU models use airflow (AF) and relative humidity (RH) features to guess what the temperature will be inside (T). In Stage 2, RF models employ these anticipated temperatures, AF, and RH characteristics to figure out how much energy the HVAC system uses overall (WHRTUTotal). The results indicate that the integrated GRU + RF framework surpasses individual models, attaining enhanced predictive precision (R² = 0.886, RMSE = 755.06) while concurrently yielding significant energy conservation (13.82% utilizing predicted T and 3.41% employing actual T). These outcomes underscore the dual advantages of precision and energy efficiency, positing that hybrid models may furnish more dependable instruments for the management of energy in intelligent buildings. The recommended methodology offers a scalable basis for incorporation into real-time control strategies like Model Predictive Control (MPC).